Welcome back to the weekly AI news roundup.

Search all 95 Sunday Bonuses

In case you missed it, here’s this week’s Thursday deep dive:

🗞️ AI news

Here’s what happened in AI this week:

👩‍💻 AI releases

1. Alibaba open-sourced Qwen3.6-27B, a cost-efficient 27B model that outperforms much larger ones on agentic coding tasks.

2. Anthropic news:

   1. Claude Connectors expanded to 200+ everyday apps, letting Claude pull in relevant data mid-chat and chain multiple services together.

   2. Cowork can now build Live Artifacts: dashboards and trackers connected to your data sources that auto-refresh when opened.

3. DeepSeek open-sourced DeepSeek-V4, a reasoning model with a 1M-token context window that approaches the performance of top-tier closed models.

4. Google news:

   1. AI Studio now gives Google AI Pro and Ultra subscribers higher usage limits and new models for prototyping and vibe-coding projects.

   2. Deep Research and Deep Research Max can perform autonomous research across your files or specialized data sources using MCP connections.

   3. Gemini for Home added a Continued Conversation feature that lets your smart speaker retain context after a response and handle follow-up questions.

5. Microsoft added Copilot’s agentic capabilities to Excel, PowerPoint, and Word, letting it perform multi-step actions directly in your documents.

6. Moonshot AI open-sourced Kimi K2.6, a coding model that can coordinate 300 simultaneous sub-agents for complex, long-horizon programming projects.

7. OpenAI news:

   1. ChatGPT for Clinicians helps verified US physicians, nurse practitioners, and pharmacists with medical research and documentation work.

   2. ChatGPT Images 2.0 is a SOTA image-generation model with thinking capabilities that can reason through your prompts.

   3. Codex Chronicle builds context from your screen activity, so Codex knows what you’re working on automatically. (Opt-in preview for Pro subscribers.)

   4. GPT-5.5 is a flagship model for complex coding, research, and data analysis that is also more token-efficient than competing models.

   5. Workspace Agents let teams build shared AI agents for long-running tasks in the cloud. (For ChatGPT for Business, Enterprise, and Edu plans.)

8. TwelveLabs launched Pegasus 1.5, which extracts structured, timestamped metadata from videos up to two hours long so you can better analyze them.

9. X launched Custom Timelines that lets you pin up to 75+ AI-curated topics to your home tab. (In early access for Premium subscribers.)

🔬 AI research

1. Alibaba announced Qwen3.6-Max-Preview, with improved agentic coding, world knowledge, and instruction following for real-world tasks.

📖 AI resources

1. “GPT 5.5, DeepSeek V4, and the Compute War” [VIDEO]: great level-headed deep-dive by AI Explained.

🤦‍♂️ AI fail of the week

Well, at least you’re aware of your own limitations, Gemini.

📹 Live & Learn #3: AI Slide Deck Makers

The “Live & Learn” format is really coming into its own, thanks to great feedback from everyone who responded to my surveys. Keep your suggestions coming!

From now on, there’s more structure to my live tests and the scoring.

Future “Live & Learn” sessions will:

• Focus on 3-5 tools in the same category for an apples-to-apples comparison

• Have 3-5 specific dimensions we rate them on

• Wrap-up with a leaderboard we fill out together during the session

This week, we tested five AI slide deck makers and ended with a two-way tie for first place:

Watch "Live & Learn #3"

Next time, we’ll be testing AI tools that automatically create helpful visuals like charts and diagrams out of text. It’s happening on Friday, May 1, at 2:00 PM CET.

Join "Live & Learn" - May 1, 2:00 PM CET

Please let me know what you want to get out of this particular session and “Live & Learn” in general:

Help shape "Live & Learn"

(It’s just two open-ended questions with a few AI-assisted follow-ups.)

Thanks!

I put five AI slide deck makers head-to-head with the same prompt.

All of them promise the same thing: describe the deck you want, hit “go,” and get a finished presentation in minutes.

But they take different routes to get there. Some want lots of feedback before you even get going; others just run with it. Some prioritize making edits via chat; others let you do it manually right away.

I ran each one through the same six-slide brief and rated them on prompt following, accuracy, design, and ease of use.

If you’ve been eyeing any of these tools but haven’t had time to try them yourself, take a look.

Tools & scores

Read more

Paid subscriber bonus: Grab the Claude Code Essentials pack with self-building skills, customizable workflows, and copy-paste use cases.

If you’re like me…you may well be my long-lost twin. If so, please reach out!

Silliness aside, if you’re like me, you’ve been building a whole bunch of skills for Claude Code to help with your daily routines.

(If you haven’t worked with skills before, check out my intro to Claude Skills.)

But I’ve come to realize something: Simply asking Claude Code to create a skill and then moving on leaves a lot of potential on the table. Skills rarely come out perfect on the first try.

So today, let’s look at how you can test and improve your Claude Code skills.

The problem with fire-and-forget skills

At its core, a skill is just a set of instructions that tell Claude what to do when the skill is invoked.

Simples!

And because the basic idea is so straightforward, you likely don’t treat your Claude skills as something to bother optimizing.

The thing is, even a skill that technically works can have a bunch of hidden issues:

• Directions that aren’t as specific as they could be

• Sub-optimal descriptions that mean the skill doesn’t trigger reliably

• Output examples that aren’t helpful for your use cases

• Redundant or overlapping instructions that confuse the agent

• Long-winded passages of text that waste tokens without improving outcomes

But as long as the output of your skill is “good enough,” you might never bother to review it and pick up on any of this.

I believe in continuously improving your Claude Code workflows.

That’s why one of the skills in my “Claude Code Essentials” hub is the Workspace Auditor, which tells you how to optimize your working folder and Claude Code setup.

And the more I learn about the topic, the more I’m convinced that any skill can be improved at least once.

Even if you rely on just one Claude skill, it’s worth figuring out how to make it better.

Luckily, there’s a beginner-friendly way to test and improve your skills.1

And it comes directly from the crew behind Claude itself, so you know it’s legit, bro.

What’s new in Anthropic’s Skill Creator 2.0?

In January, when I first wrote about this, I recommended a meta-skill from Anthropic called “Skill Creator.” It helps Claude build better skills by codifying how to structure a skill, write its instructions, create description triggers, and so on.

In early March, Anthropic rolled out an improved version of the Skill Creator. (Check it out on GitHub.)

The new version is a serious step up.

For one, it contains even more robust instructions about the purpose of a skill, what it should include, and its structure:

But beyond that, Anthropic added a suite of scripts, templates, sub-agents, and instructions that make Claude actually test your skill and improve it.

Check out the gray box for details if you’re curious.

The agents work together as a proper A/B testing framework for any Claude skill. You run two versions of a skill, compare results, pick a winner, then rinse and repeat.

Let me show you how this happens in practice.

Subscribe now

How do Skill Creator’s test runs work?

The cool part is that your role is minimal: You review test results and give Claude feedback (if any).

That’s it.

The Skill Creator and its agents do all the testing and skill optimization behind the scenes.

If you’re interested in the details, check out the grey box.

Otherwise, see you at the bottom of it. Bye!

For a hands-on showcase, I created a simple “Tagline Writer” demo skill and ran it through Skill Creator’s gauntlet above.

Here’s the “Eval Review” report I got from the Skill Creator:

Don’t worry, I don’t expect you to squint your eyes to see every detail.

I’m not a monster.

Let’s break it down.

Here’s the test prompt and output for a single “with skill” run:

Here are the checks and grades for the six assertions:

Here’s the feedback box that lets me share my own thoughts:

Here are benchmark comparisons for “skill” vs. “no skill” runs:

And finally, here are Analyzer’s top-level observations:

I get concrete outputs from every run, which is neat.

If I’m happy with the skill, I’m done.

If not, I ask the Skill Creator to optimize the skill based on the scores, its analysis, and my input, and then it runs another batch of tests and feeds me the report.

It’s a well-defined, repeatable process, and the Skill Creator runs it right out of the box.

BUT!

I was honestly quite surprised to find that a key element of the process is basically treated as an afterthought.

The one major thing Anthropic overlooked…

Remember those “assertions” I mentioned? The ones that determine if your skill passes or fails?

They’re the single most important thing to nail in the entire process.

Every “your skill improved by 40%” stat is only as good as the quality of these assertions.

As such, you’d think Anthropic would have robust definitions for how to write these and a set of checks to make sure they’re actually relevant to the skill in question.

Instead, assertions get two short paragraphs of vague instructions:

Don’t just wait for the runs to finish — you can use this time productively. Draft quantitative assertions for each test case and explain them to the user. If assertions already exist […] review them and explain what they check.

Good assertions are objectively verifiable and have descriptive names — they should read clearly in the benchmark viewer so someone glancing at the results immediately understands what each one checks. Subjective skills (writing style, design quality) are better evaluated qualitatively — don’t force assertions onto things that need human judgment.

That’s literally it.

“Draft some assertions, make them verifiable, use your judgment.”

The Skill Creator has three dedicated agents, several standalone scripts, structured templates, and lengthy instructions for running the tests, but leaves the actual eval metrics up to Claude’s intuition.

That’s straight-up crazy to me!

It’s as if Anthropic built the world’s most sophisticated car but told you to fill it up with, “like, whatever, some fuel or something, I dunno lolz.”

Jokes aside, this matters more than you’d think.

If the assertions Claude writes are crap, the whole test engine spins up and eats your token quota only to conclude that “the output has letters in it, so that’s a ‘pass,’ my man.”

As far as I’m concerned, a good assertion or eval metric must:

• Be grounded in what the tested skill promises to do (its primary purpose)

• Check for not just “does it do the thing?” but “does it do the thing well?” and also “does it avoid doing the wrong thing?”

• Be specific enough that the Grader can evaluate it unambiguously

And for some reason, the Skill Creator tells Claude to figure all of this out on the fly, in the background, while the test agents are already running, as if it’s a fun little side project to pass the time.

This really bugged me, so I went and did something about it.

Subscribe now

My “Eval Maker” skill and what it does

To patch this gap, I created the Eval Maker.

Eval Maker analyzes the Claude skill you want to improve and surfaces its primary job: What does the skill promise to do, both explicitly and implicitly?

If the skill’s purpose connects to established best practices, Eval Maker will identify real-world frameworks you can use to evaluate its quality.

Only then, based on this analysis, does Eval Maker design evals (assertions) for the Skill Creator to use in its test runs.

You get a nice interactive HTML report with the following:

• An overview of your skill and what’s already working

• Quick fixes you may want to implement before running full Skill Creator tests

• Three recommended test prompts for the Skill Creator to use

• A set of recommended high-impact assertions for the Skill Creator (plus a list of optional ones you can include)

• A single copy-paste prompt that invokes the Skill Creator and feeds it everything needed to kick off and run its tests

Here’s a sneak peek at how this looks for a simple demo skill called “Tweet Trimmer”:

1. Your skill at a glance

2. Test prompts

3. Checks to run (evals/assertions)

4. Copy-paste prompt

Here’s a good way to think about it:

• Anthropic’s Skill Creator gives you the dashboard to measure your skill.

• Eval Maker tells you what’s actually worth measuring and optimizing for.

Here’s how to get both of them working together in a few minutes:

Read more

Welcome back to the weekly AI news roundup.

Search all 95 Sunday Bonuses

In case you missed it, here’s this week’s Thursday deep dive:

🗞️ AI news

Here’s what happened in AI this week:

👩‍💻 AI releases

1. Alibaba open-sourced Qwen3.6-35B, a sparse coding agent that matches larger models on agentic tasks. (Try for free on Hugging Face.)

2. Anthropic news:

   1. Claude Code “Routines” let you schedule automated workflows that run in the cloud without relying on your computer being awake.

   2. Claude Desktop app was redesigned for better Claude Code workflows with parallel coding agents, built-in terminal, in-app editing, and more

   3. Claude Opus 4.7 is the newest frontier model with improved reasoning and complex agentic coding, as well as better vision capabilities.

   4. Claude Design turns prompts into polished prototypes, slides, and visuals with built-in collaboration and design elements.

3. Blackmagic Design released DaVinci Resolve 21 with a dedicated Photo page and expanded AI editing tools for video and image workflows.

4. Character.ai news:

   1. PipSqueak 2 is a faster and smarter chat model paired with an upgraded Memory, so your characters remember more details across conversations.

   2. c.ai Books lets you “step inside” classic novels like Alice in Wonderland and role-play through the plot as different characters.

5. Google news:

   1. AI Studio added a Tab Tab Tab prompt autocomplete feature to its vibe-coding environment, so Gemini can intelligently supplement your ideas.

   2. Chrome added Skills that let you save and trigger custom Gemini prompts with one click directly from the browser.

   3. Gemini app is now available on Mac as a native desktop experience.

   4. Gemini 3.1 Flash TTS is the latest text-to-speech model with audio tags that let developers build AI voice apps with more natural, expressive speech.

   5. Google desktop app for Windows has built-in AI Mode search and lets you share your screen to ask questions about its contents.

   6. Personal Intelligence uses Nano banana to generate AI visuals from your Google Photos library.

6. Microsoft launched MAI-Image-2-Efficient, an image model that’s 22% faster and 41% cheaper than its bigger cousin.

7. Midjourney launched V8.1 Alpha, which generates native 2K HD images three times faster and cheaper than V8 and features an overhauled “Describe” tool.

8. Mozilla released Thunderbolt, an open-source, self-hosted AI client that keeps your conversations local and private.

9. OpenAI updated Codex with computer use, a built-in browser, 90+ plugins, image generation, and memory, turning it into a full-on desktop agent.

10. Opera launched a Browser Connector MCP that lets Claude and ChatGPT read your open tabs and take screenshots directly in the browser.

11. Perplexity released Personal Computer for Mac that can manage local files, apps, and web actions in a secure sandbox environment. (Max subscribers only.)

🔬 AI research

1. Adobe previewed Firefly AI Assistant, a tool that automates multi-step tasks across Creative Cloud apps in response to plan-language prompts.

2. Canva announced AI 2.0, a major platform overhaul with prompt-based editing, persistent brand memory, and HTML import.

3. Microsoft is testing OpenClaw-style always-on AI agents inside Microsoft 365 Copilot that proactively handle business tasks across apps.

📖 AI resources

1. “2026 AI Index” [REPORT]: Stanford HAI’s annual deep-dive on the state of AI with key takeways on adoption trends, economic impact, etc.

2. “Automated Alignment Researchers” [REPORT]: Anthropic’s study on using Claude to automate scalable AI oversight research.

3. “Half of Americans Use AI Services” [POLL]: Ipsos/Epoch poll of Americans’ AI usage and workflows.

🤦‍♂️ AI fail of the week

When Gemini thinks you’re the epitome of classic beauty…

📹 Live & Learn #2: Four AI app builders

Last session, I tested four no-code AI app builders with the same app request. The contenders came in different flavors and had their own spin on the task.

Watch "Live & Learn" #2

During the next “Live & Learn,” I’ll be test-driving several AI slide deck makers.

The session will be on Friday, April 24, at 2:00 PM CET:

Join Live & Learn - April 24, 2PM CET

To help me make the session more useful and relevant, share your thoughts:

Take The Survey

(The survey is just one open-ended question with AI-assisted follow-ups.)

Thanks!

All of them have the same premise: You describe what you want them to build, hit “go,” and get a working prototype in minutes.

But they come in different flavors: Some are chat-based, some cobble a flow together from preset blocks, and some vibe code new interfaces and apps from scratch.

If you’re curious about these tools but haven’t had a chance to check them out yourself, this is a good starting point:

Tools tested & app prompt:

Read more

EDIT (April 15, 2026): Article updated with five new Gemini tests.

In March, Anthropic and OpenAI both released interactive visual explainers for their chatbots. Google followed suit a month later.

These explainers supplement your text chat with diagrams, charts, and other visuals you can manipulate to help you better understand concepts.

This brought me right back to August 2025, when the same three companies near-simultaneously released their versions of “study” modes:

Just as I did then, I wanted to see how the ChatGPT, Claude, and Gemini versions compared.

So I went ahead and tested them, because who’s gonna stop me?!

How do the interactive explainers work?

Although the idea is similar, the three companies approach it differently.

ChatGPT

OpenAI chose to curate a shortlist of 70+ STEM concepts in advance, so when you ask about them, ChatGPT should automatically pull up the relevant visual explainer.

These explainers are pre-built and will always look and work the same way.

Claude

Claude, instead, designs a new interactive explainer from scratch every time:

Because of this, Claude’s explainers aren’t limited to math and science topics and can also be requested on demand with relevant commands:

Claude will decide when to build a visual for something, or you can ask it to do so directly (with a query like “draw this as a diagram” or “visualize how this might change over time”).

Gemini

Gemini works a lot like Claude, and it tends to provide both the in-depth text explainer and an optional “Show me the visualization” button by default:

Here’s Google’s example with Gemini visualizing how fractals work:

So how does this work in practice?

Let’s find out!

The five tests

Since ChatGPT only has its preset library to pull from, this can’t be a true apples-to-apples comparison.

Instead, I’ll kick off with three explainers from ChatGPT’s list and supplement with two freeform ones that Claude and Gemini might be better able to build on the fly.

Test #1: Pythagorean theorem

Let’s take a trip down memory lane back to our school years and basic geometry.

Prompt: “Explain the Pythagorean theorem.”

ChatGPT

ChatGPT’s version is pretty barebones, but it does the trick:

I can adjust the lengths of the a and b sides (legs) to see how this affects the c side (hypotenuse). The formula below the sliders explains the relationship.

Claude

Claude didn’t trigger any visuals automatically and provided a text explainer first:

But when prodded to “Show me interactively,” Claude built what’s arguably a better version of ChatGPT’s pre-made diagram:

I like that Claude color-coded the sides and mapped the calculations directly onto the corresponding squares. Much easier to connect the dots!

But I do wish Claude had included the square root calculation for the c side to really bring this home.

Claude also lets me save the resulting diagrams as Claude Artifacts, so I’ll be sharing them here:

Check out the Artifact

Gemini

Unlike ChatGPT, Gemini also built the visualization from scratch.

And unlike Claude, Gemini offered to visualize it without being explicitly asked to:

The result is very similar to the other two models, but Gemini made questionable decisions about text color and placement for the “leg” labels. It also doesn’t explain what’s happening quite as well as Claude does.

My take

While all three do largely the same thing, I find that Claude’s version is the easiest to parse at a glance, even though it doesn’t provide the final connection by visually mapping the square root calculation back to the length of the hypotenuse.

Test #2: Mirror equation

Yet another equation I’d long forgotten everything about. Fun!

Prompt: “Show me the mirror equation.”

ChatGPT

ChatGPT uses the same clean blue-line diagram style for this:

Claude

Claude made a visual by default this time, but a static rather than interactive one:

Again, this was an easy “Show me interactively” fix:

While Claude’s visual is again prettier to look at and shows helpful color-coded effects, the lines of the diagram don’t always connect to the mirror in a way that’s easy to grasp:

Check out the Artifact

Gemini

This time, I had to nudge Gemini to create a visualization, but it ended up with a similar concept as the other two models:

But it suffered from even more visual issues than Claude did. The mirror itself would slide around instead of staying fixed, which certainly doesn’t help follow the logic.

My take

I’m still impressed by Claude’s ability to create on-the-fly visuals that are very close to what ChatGPT has pre-coded for it. Bonus points for letting users pick and visualize concave vs. convex mirrors.

The downside is that diagram elements sometimes connect in ways that aren’t fully clear, so the value you get from them depends on your existing understanding of the concepts.

The problem is even more pronounced with Gemini. Even though it provides the right sliders and selections, the movement of objects is too inconsistent for the visualization to be truly helpful.

Test #3: Ideal gas law

If you’d asked me to explain this one before working on the article, I’d have stared at you with a blank expression before slowly backing away. But here we are.

Prompt: “Demonstrate the ideal gas law for me.”

ChatGPT

I like that the animation makes it instantly clear how gas molecules speed up at higher temperatures and why pressure might increase as the volume decreases. It’s perhaps the most visually intuitive of ChatGPT’s examples.

Claude

Claude made a graph with slider options instead of a visualized container:

But I wanted to see if Claude could mimic ChatGPT’s version, and it 100% could:

You get the same intuitive feel for molecule speed and pressure impacts, with better layman-friendly labels, too.

Check out the Artifact

Gemini

Gemini triggered the visualization option by default and ended up with something very similar to Claude again:

I like the subtle touch where the moles change color at different temperatures, and I find the shrinking container visual better than Claude’s.

My take

The pattern is clear: Claude can usually match ChatGPT’s visuals, but it often needs a bit of prompting to get there, which requires the person to know what they’re after. ChatGPT’s visuals should pop up by default to supplement its text-based answers.

I prefer Gemini’s interactive visual to Claude’s this time around. It also required less additional prompting to build the visualization.

Test #4: Combustion engines

We’re now moving out of ChatGPT’s pre-programmed “comfort zone.”

Prompt: “Show me how an internal combustion engine works.”

ChatGPT

Oh man, this was a quadruple fail right out of the gate.

ChatGPT instantly defaulted to text, ignoring the “show” qualifier completely:

When I nudged it using the “show me interactively” (which worked for Claude), ChatGPT created an image:

I then tried to be even more explicit by saying, “I want something I can actually manipulate and interact with.”

ChatGPT gave me a text buffet of options. Finally, I had to ask it for HTML outright:

The result was super basic and didn’t do much to explain the individual steps:

No helpful labels of any kind. The piston escapes outside the cylinder. Manipulating speed doesn’t affect the way the engine works, so the decision to include it is questionable.

Finally, ChatGPT couldn’t create anything shareable, so I had to paste its HTML into a third-party site to be able to share it here:

Check out ChatGPT's "explainer"

Claude

For once, Claude nailed the task on the very first try:

It spit out a clean, visually pleasing four-step diagram that I could click through to understand the concept clearly.

Take a look for yourself:

Check out the Artifact

Gemini

Gemini followed in Claude’s footsteps yet again, but also added a neat animation feature to show the process in action.

Is it just me, or, when animated, the steps don’t match up with the motion of the piston? When clicking on the static steps, everything’s correct, but as soon as the engine starts moving, the piston ends up in the opposite spot at every step.

My take

Man, this wasn’t even close.

Not only did Claude intuitively know what I needed, but the result was significantly better with far less effort.

Gemini is a close second and could have even outshone Claude with its animation feature, but unfortunately, the animation did more harm than good in this case.

Test #5: Tectonic plates

Let’s give ChatGPT a chance to redeem itself!

Prompt: Show me the major tectonic plates and their movements.

ChatGPT

Oh no, here we go again:

ChatGPT has apparently never heard of “show, don’t tell.”

Let’s be more explicit again:

Behold, the “Tectonic Plate Map”:

There you have it, kids: Tectonic plates are unidentified colorful rectangles that are permanently stuck to each other!

Any questions, class?

Claude

Again, Claude did better right out the gate:

But I wanted this to be more interactive, so I nudged:

It worked!

The visual isn’t winning any awards for aesthetics or geography, but it does actively aid my conceptual understanding and provides useful info about each plate in a clean modal box.

The list of major plates is complete and accurate, according to Wikipedia.

Check out the Artifact

Gemini

Gemini again ends up with something similar to Claude:

And again, there are caveats. Despite listing the major tectonic plates in its text response, Google decided to visualize the minor Nazca plate and failed to show the Antarctic plate, which is a major one. Also, its directional errors different a lot from Claudes, so I had to double check. Claude got more of them correct based on my research.

Finally, I’m not sure what the benefits of extending the arrrows or switching to a “Physical Map” are, since the visualization is abstract anyways.

My take

Claude is consistently more visually engaging and genuinely helpful at showcasing concepts on the fly. Gemini is generally similar but falls short of both Claude’s look and accuracy.

ChatGPT struggles to visualize anything outside of its library of approved templates. Even with repeated nudging, the outcome isn’t nearly as polished as Claude’s versions.

Subscribe now

General observations

Here are my concluding thoughts.

Note that they’re based exclusively on my limited tests with free versions of each tool.

I’d expect paid model options like GPT 5.4 Thinking and Opus 4.6 to handle these tasks better.

Google does let you its Gemini 3.1 Pro version for free (with a strict quota), so that’s what I’ve been using.1

However, since interactive explainers are marketed as being available to everyone, I think it’s only fair to mimic the experience of the average user.

ChatGPT (“The Curator”)

While it didn’t work for me, the idea behind human-approved visualizations popping up automatically for given topics is solid.

The good: For the pre-approved topics, ChatGPT should load the visuals almost instantly. They’ll be precise, predictable, and guaranteed to show accurate equations. Also, since they trigger automatically, the user doesn’t even have to know this option exists or learn how to invoke it.

The not-so-good: The visuals are rather basic and often feel dry and technical. When you move beyond the shortlist, ChatGPT truly struggles to create helpful, interactive elements on its own.

Who this is for: School and high school students who have specific STEM concepts fresh in their minds and need to anchor them with visuals.

Claude (“The Designer”)

Despite not having a curated library of concepts, Claude is strictly better at coding polished interactive elements from scratch, even on a free account.

The good: Claude’s explainers almost always look pretty, have more helpful labels, and—at least in my tests—are easier to grasp at a glance, especially fo concepts that don’t come from ChatGPT’s STEM list.

The not-so-good: Since Claude’s explainers are always designed on the fly, they’re less predictable. You’re essentially rolling the dice whenever you ask for a new visual. Without a robust review, Claude’s outputs are subject to the usual AI hallucinations, which might defeat the purpose of helping a layman understand brand-new concepts. Finally, Claude often needs prodding to create an interactive element, so people who aren’t aware of this feature might not even get to experience it at all.

Who this is for: People who want to grasp almost any complex concept in a flexible, visual, and interactive way.

Gemini (“The Animator”)

Gemini usually provides a hybrid answer that explains the concept in writing while also providing an optional “Show me the visualization” step.

The good: Gemini reliably triggers visualizations for most learning topics without being prompted. This is more likely to expose the “interactive explainer” feature to people who aren’t already aware of it. Gemini is also eager to try animating the visuals.

The not-so-good: While Gemin’s conceptual explainers end up similar to Claude’s, they’re typically less polished and more buggy, at least in my testing. In some cases, this undermines the entire purpose of an explainer.

Who this is for: People who don’t know the interactive option exists and want to get a feel for its potential for the first time.

🫵 Over to you…

Of course, these are just my subjective thoughts based on a limited range of tests.

You may get very different results.

So take these for a spin yourself and let me know what you think.

Pick a concept you always struggled with and see if interactive explainers make it stick!

Thanks for reading!

If you enjoyed this, here’s how you can help:

• ❤️Like this post if it resonates with you.

• 🔄Share it to help others discover this newsletter.

• 🗣️Comment below—I love hearing your opinions.

Why Try AI is a passion project, and I’m grateful to those who help keep it going. If you’d like to support me and unlock cool perks, consider a paid subscription:

Subscribe now

Welcome back to the weekly AI news roundup.

Search all 95 Sunday Bonuses

In case you missed it, here’s this week’s Thursday deep dive:

I was away during Easter, so we have two weeks’ worth of AI developments to catch up on. Let’s get to it!

👩‍💻 AI releases

1. Alibaba news:

   1. HappyHorse-1.0 is a SOTA video model that anonymously topped global video leaderboards before being claimed by the company.

   2. Qwen3.5-Omni is a family of natively omnimodal models that can simultaneously process and generate text, images, audio, and video.

   3. Qwen3.6-Plus is the upgraded flagship model with stronger reasoning and coding capabilities that can reliably handle complex real-world tasks.

2. Anthropic news:

   1. Claude Cowork is now generally available with admin controls like role-based access, spend limits, and analytics for company-wide deployment.

   2. Computer Use now runs directly in Claude Code CLI, so it can visually verify and fix what it builds by interacting with the screen.

   3. Microsoft 365 Connectors rolled out for all Claude plans, so you can link Claude to your OneDrive, Outlook, and SharePoint directly.

3. Arcee AI open-sourced Trinity-Large-Thinking, a reasoning model scoring #2 after Cladude Opus-4.6 on PinchBench while being 96% cheaper. (Try for free.)

4. Cursor now lets you launch coding agents from your phone and run and control them remotely on any machine.

5. GitHub added a focused review agent called “Rubber Duck” to Copilot CLI that uses a second model to review code and catch errors before they compound.

6. Google news:

   1. Gemini app added interactive simulations that showcase concepts right in the chat. (I previously compared ChatGPT and Claude versions.)

   2. Gemini added Notebooks for organizing chats, files, and custom instructions per project, with built-in syncing to NotebookLM.

   3. Gemma 4 is a family of open models for advanced reasoning and complex multi-step tasks that is faster and cheaper to run on edge devices.

   4. Google AI Edge Eloquent is a free offline dictation app for iOS that transcribes speech and removes filler words automatically. (Download it here.)

   5. Google Colab introduced a Learn Mode and Custom Instructions, turning Gemini into a personalized coding tutor.

   6. Google Vids added Veo 3.1 video generation with 10 free monthly clips, plus AI avatars and custom music tools.

   7. Veo 3.1 Lite is a cost-efficient video model that’s more than 50% cheaper than Veo 3.1 Fast.

7. HeyGen news:

   1. Avatar V can capture and consistently recreate your likeness in any setting, from just a 15-second video recording.

   2. Topaz Labs is now integrated into the platform, letting you natively upscale videos to 4K.

8. Meta is rolling out a new flagship model called Muse Spark, capable of multimodal reasoning and visual understanding across all Meta AI touchpoints.

9. Microsoft news:

   1. MAI-Transcribe-1 turns speech into text across 25 languages, even in noisy environments, and is both faster and cheaper than existing offerings.

   2. MAI-Voice-1 generates expressive speech from just a few seconds of input and can produce a minute of audio in under a second.

10. OpenAI news:

    1. ChatGPT Pro now has a $100/month tier with 5x more Codex usage than Plus.

    2. Codex now offers pay-as-you-go seats for teams with reduced annual pricing for ChatGPT Business.

    3. Codex launched as a plugin for Claude Code, so you can use OpenAI’s coding models inside Claude Code workflows.

11. Perplexity news:

    1. Computer launched inside Slack to help teams run analysis, generate code, and trigger workflows directly from Slack channels and DMs.

    2. Computer for Taxes uses IRS materials to help US users prepare their federal returns through its AI agent.

    3. Plaid integration lets you securely link bank accounts, credit cards, and loans for a full view of your personal finances.

12. Pika launched a Video Chat Skill (in beta) powered by PikaStream 1.0 that gives any AI agent a face and voice for real-time video conversations.

13. Salesforce gave its Slackbot 30+ new features, including automated meeting notes, daily recaps, and AI-powered task management.

14. Softr launched a no-code platform with an AI co-builder that creates functioning business apps from plain-language descriptions.

15. Topaz Labs shipped a Precision update, making its video models faster and better at upscaling clips to 4K.

16. World Labs rolled out Marble 1.1 and Marble 1.1 Plus, with improved visual details and larger, more complex generated worlds.

17. X added a Grok-powered photo editor to the post composer, letting you ask for edits and tweaks to images before posting.

18. Z.ai open-sourced GLM-5.1, a coding model that can work autonomously for up to eight hours and tops several software engineering benchmarks.

🔬 AI research

1. Anthropic’s Claude Mythos Preview is a powerful ureleased frontier model capable of finding and exploiting previously hidden software vulnerabilities.

📖 AI resources

1. “Claude Mythos: Highlights from 244-page Release” [VIDEO]: An excellent deep dive by AI Explained.

2. “Emotion concepts and their function in a large language model” [RESEARCH]—study by Anthropic on distinct emotions that shape how a model behaves.

🔀 AI random

1. Anthropic accidentally leaked the Claude Code source files, exposing internal architecture and planned features and sparking security concerns.

🤦‍♂️ AI fail of the week

When a vectorized version of you ends up looking like knock-off Freddie Mercury.

📹 Live & Learn #1: Testing AI tools together

In the first-ever “Live & Learn” session, I learned a tiny bit about AI tools and a whole lot about Substack’s screen-sharing limitations and the importance of double-checking audience settings. (I explain more here.)

Watch the recording

I really want these “Live & Learn” sessions to become a fun way for us all to nerd out about AI together. They don’t have to be limited to the AI tools I prepare for testing.

So bring your own tool suggestions, questions, and other AI things to share with me and the audience. I’ll make sure the live chat is actually open to all subscribers next time. (Oops!)

Join "Live&Learn" - April 16, 2:30PM CET

Also, please let me know what you’d like to get out of these live sessions:

Help shape "Live & Learn"

(It’s just a single open-ended question with a few AI-assisted follow-ups.)

Thanks!

1. Create scalable vector graphics SVGs from text prompts or uploaded images.

2. Intelligently add music and sound effects to uploaded silent video clips

3. Act as your AI co-writer, making trackable changes, doing research, etc.

And speaking of living and learning. This session taught me that:

1. I should pay attention to Substack’s “audience” and “comments” settings: This session was apparently limited to existing subscribers rather than available to everyone, and only paid subscribers were allowed to post comments.

2. Screen sharing on Substack Live does not share tab audio, so I’ll have to limit myself to tools that don’t rely on sound effects.

3. One must always double-check the mic, as I apparently went silent for the last three minutes of the stream.

Live and learn!

The three tools I tested:

Read more

I hope you had a blast!

I’ve done a bit of soul-searching recently, which is glorified influencer-speak for “I figured I’d try some new stuff.”

As a result of that searching of the soul, I’ll be making two adjustments to the newsletter.

Let me tell you what I’m up to and how it CHANGES EVERYTHING might slightly affect your inbox on Thursdays and Sundays.

1. “Sunday Bonus” → After-post bonuses

The first change is all about regular paid perks.

What I’ve been doing so far:

It’s hard to believe, but the very first Sunday Bonus1 went out over two years ago, on March 3, 2024.

Today, we’re at exactly 95 Sunday Bonuses in total. If my math checks out, that’s almost 100 bonuses!

I love throwing extra perks at paid subscribers.

But in retrospect, having to come up with a bonus every Sunday was a bit of an own-goal situation.

First, on some weeks, these bonuses honestly felt somewhat forced and rushed, which didn’t do wonders for their quality.

Second, Sunday Bonuses were often related to topics I explored earlier on Thursdays. Running them separately created an unnecessary disconnect between the stories.

Finally, the self-enforced obligation to deliver a Sunday Bonus every week meant I didn’t always dedicate enough time to a Thursday post to do it justice. There are only so many hours in a day.2

What I’ll start doing instead:

From now on, whenever I have genuinely useful add-ons for paid subscribers—tools, templates, Claude Code skills, etc.—I’ll share these in a paywalled section of the corresponding Thursday post.

This gives me the flexibility to keep certain Thursday posts free for everyone while also having more time to work on actually valuable paid add-ons for the rest of them.

Sunday Rundowns with AI news will continue, and all existing Sunday Bonuses will stay as they are.

Subscribe now

2. New “Live & Learn” sessions

And now for something a little different…

What I’ve been doing so far:

I’ve held several live workshops over the years, but it’s mostly been an ad hoc kind of thing.

Also, if I’m brutally honest, painstakingly preparing properly polished presentations3 is not really my jam.

What got me into writing about AI in the first place was the joy of discovery, not holding “scale your business and 10x your productivity with the power of automated agentic workflows” seminars.

Luckily, I have a plan to bring back that spirit of open-ended exploration and take you along for the ride.

What I’ll start doing instead:

Since late February, I’ve been having weekly unscripted Substack Live chats with Leor of . They’re lots of fun, and I am especially enjoying their open-ended, off-the-cuff nature.

On top of that, Substack recently launched a native screensharing feature, so I can now show you my desktop in realtime without hacky third-party streaming platform shenanigans.

So every week (starting tomorrow), I’ll be going live on Substack to check out random AI tools, poke at their rough edges, take audience requests, and just have fun testing what’s out there.

If that’s your cup of tea, join me:

Join "Live & Learn" - April 10, 2PM CET

Paid subscribers who aren’t able to join live will get session recordings afterwards.

🫵 Over to you…

What do you think of these upcoming changes? Good? Bad? Neutral?

Do you have other suggestions to make Why Try AI more relevant and useful to you?

I’m all ears!

Leave a comment

Thanks for reading!

If you enjoy my work, here’s how you can help:

• ❤️Like this post if it resonates with you.

• 🔄Share it to help others discover this newsletter.

• 🗣️Comment below—I love hearing your opinions.

Why Try AI is a passion project, and I’m grateful to those who help keep it going. If you’d like to support me and unlock cool perks, consider a paid subscription:

Subscribe now

As usual, I updated my list of go-to AI tools to bring you all up to speed:

Enjoy,
Daniel

LATEST UPDATE: April 1, 2026

I first published this round-up on October 17, 2024. But things in AI move fast, so I now release regular updates. Check out this archive page with all the past versions (fun for tracing how AI and my preferences evolved).

I often get asked about AI tools that I personally use.

While I regularly test and review all sorts of AI models and sites, most don’t end up a part of my routine.

Here are those that did.

Subscribe now

🗣️ 1. Large language models

These are the chatbots and LLMs I turn to.

• 🎁Free:

  • Claude Sonnet 4.6 [UPDATED - replaced 4.5] for most creative work. Claude has a “fresh” conversational feel and tends to suggest unconventional ideas. The new interactive explainers are awesome for visualizing and grasping tricky concepts. The only downside is the strict message limit for free accounts.

  • Gemini 3.1 Pro (via Google AI Studio) [UPDATED - replaced 3.0] for working with big files. Gemini’s massive token window and native processing of audio and video make it a beast for handling long inputs in multiple formats.

  • GPT 5.4 [UPDATED - replaced GPT 5.2] still holds up for regular day-to-day chats and low-stakes research, although I find myself gravitating away from ChatGPT toward alternatives.

• 💵Paid:

  • Claude 4.6 Opus as my primary daily driver for everything I do with Claude Code. It runs and organizes most of my work and newsletter projects.

🖼️ 2. AI images

Here are my favorite image models.

• 🎁Free:

  • Gemini 3.1 Flash Image (aka Nano Banana 2) [UPDATED - replacing Nano Banana Pro] for image editing. Excellent at detail-preserving image edits from natural-language requests.

  • GPT Image 1.5: Still the top-ranked image generation model that understands context and can render entire pages of text in an image. You get a free quota on ChatGPT and can also run it for free on copilot.microsoft.com.

• 💵Paid:

  • Midjourney: I’ve written dozens of posts about Midjourney over the years. It’s become harder to recommend these days, as many free models have overtaken it in terms of prompt adherence and text rendering, but it remains great for artistic exploration and photographic images.

📽️ 3. AI video

The AI video space is making the biggest leaps lately. Here’s my take on AI video models based on thorough tests of both text-to-video and image-to-video.

• 🎁Free:

  • Seedance 2.0 (via CapCut): It’s currently the best video model, and you get free daily credits (but only enough for about 1-2 clips, so use sparingly).

  • Veo 3.1 is a top-tier model for lifelike videos with native audio effects and speech. The Flow platform now gives you free daily credits to try it.

• 💵 Paid:

  • Kling 3.0 Omni: Kling isn’t quite as generous with free credits anymore, but if you’re willing to pay, 3.0 Omni is a top-tier model.

  • Any of the above: If you’re all in on either Seedance 2.0 or Veo 3.1, you can always upgrade for more generations.

🎵 4. AI music

This space is surprisingly steady. I rarely use AI music tools, but when I do…

• 🎁Free:

  • Producer for solid outputs that stick close to your prompt.

  • Suno for more creative but natural-sounding results.

Both give you a generous amount of free credits, so take them for a spin to see which one you prefer. My recent head-to-head test had me narrow my choices to the two models above (sorry, Udio and Sonauto).

• 💵 Paid:

  • Any of the above: Once you’ve picked your favorite, you can always upgrade for additional features.

🔬 5. Research

• 🎁Free:

  • NotebookLM remains a super easy recommendation. I’ve been a fan since March 2024. (I covered everything in my recent live video walkthrough.)

  • Learn About is another Google product that complements NotebookLM nicely. While NotebookLM is grounded in your preselected sources and synthesizes info across them, Learn About can browse the web and is built for open-ended exploration of any topic.

  • Perplexity I don’t use Perplexity very often, but its “Pro” search is a good midway point between simple web browsing and the time-intensive “Deep Research” tools.

• 💵 Paid:

  • Perplexity Deep Research was recently upgraded with the latest Claude models to reach SOTA performance on deep research benchmarks.

📈 6. Productivity

• 🎁Free:

  • Google AI Studio: Google still offers a whole lot for free here. You can try its latest frontier models, voice generation, coding tools, and live calls with screen or camera sharing. It’s a great way to sample most of what Google has to offer at no charge.

  • Napkin automatically turns walls of text into catchy visuals like diagrams, infographics, etc. For design-challenged people like myself, it’s a superb way to illustrate concepts with minimal effort.

• 💵 Paid:

  • Claude Code: My primary daily driver for actually building and doing stuff with AI. (Here’s just some of what Claude Code can do for you.)

  • Genspark Super Agent: A robust agent that can orchestrate dozens of tools to complete complex, multi-step tasks. You get free monthly credits to try it, but you’ll likely need the paid plan for bigger tasks.

Subscribe now

🫵 Over to you…

What AI tools do you use the most? What would you add to the above list?

Leave a comment or drop me an email at whytryai@substack.com.

Leave a comment

Thanks for reading!

If you enjoy my writing, here’s how you can help:

• ❤️Like this post if it resonates with you.

• 🔗Share it to help others discover this newsletter.

• 🗩 Comment below—I read and respond to all of them.

Why Try AI is a passion project, and I’m grateful to everyone who helps keep it going. If you’d like to support my work and unlock cool perks, consider a paid subscription:

Subscribe now

Heads up: I’m away for Easter, so the next Sunday Rundown is on April 12

Happy Sunday, friends!

Welcome back to the weekly look at generative AI that covers the following:

• Sunday Rundown (free): this week’s AI news + a fun AI fail.

• Sunday Bonus (paid): an exclusive segment for my paid subscribers.

Search 90+ Sunday Bonuses

In case you missed it, here’s this week’s Thursday deep dive:

If you’re consistently missing out on my emails, remember to check your “Promotions” tab and mark whytryai@substack.com as a “Safe Sender.”

Let’s get to it.

🗞️ AI news

Here are this week’s AI developments.

👩‍💻 AI releases

1. Ai2 open-sourced MolmoWeb, a browsing agent that navigates websites and reads the screen like a human to complete tasks. (Try the demo.)

2. Anthropic news:

   1. Claude mobile now lets you access tools like Amplitude, Canva, and Figma directly from your phone instead of relying on a laptop.

   2. Claude Code added Auto Mode that lets it run longer tasks with fewer interruptions by auto-approving safe actions while blocking risky ones.

3. ByteDance news:

   1. CapCut launched a Timeline-Free Video Studio that lets you create and edit AI video clips using plain language instead of traditional timeline editing.

   2. Seedance 2.0 video model finally rolled out to the wider public via CapCut.

4. Cohere open-sourced Transcribe, a SOTA model that accurately converts audio into text and offers multilingual support for real-world use cases.

5. Google news:

   1. Gemini now lets you import memory and chat history from other AI tools, making it easier to switch without losing your personalization.

   2. Gemini 3.1 Flash Live is a real-time voice model that delivers faster, more natural conversations and is especially well-suited for low-latency agents.

   3. Lyria 3 Pro can now generate full 3-minute tracks and parse structured elements like “verse,” “chorus,” and “bridge” prompts.

   4. Search Live expanded globally and is now powered by Gemini 3.1 Flash Live for more natural voice conversations directly inside Google Search.

6. Instagram added AI transitions to Stories that can turn a collection of still photos into an animated video clip.

7. Lovart launched a Move Object tool that lets you reposition any element in an image by selecting, dragging, and prompting for changes.

8. Luma Labs released its Uni-1 reasoning image model that thinks through your request before generating an image to ensure better prompt adherence.

9. Mistral AI launched Voxtral, a low-latency multilingual text-to-speech model that supports nine languages and lets you fine-tune the voices.

10. OpenAI upgraded ChatGPT shopping with visual browsing and side-by-side comparisons that help you find and evaluate products in seconds.

11. Smallest.ai launched Lightning V3, a text-to-speech model that supports 15 languages and lets you customize tone, pace, and style in natural language.

12. Suno released v5.5 with Voices, Custom Models, and My Taste features to create personalized music that better reflects your style and even uses your own voice.

🔬 AI research

1. Anthropic is rolling out Computer Use in Claude Cowork that lets Claude control your desktop to complete tasks (In research preview for paid accounts on macOS).

📖 AI resources

1. “ARC-AGI-3” [BENCHMARK]: new benchmark from the ARC Prize Foundation to test agent performance on previously unseen problems. (Try for yourself.)

🔀 AI random

1. OpenAI is shutting down Sora entirely to focus on its core business. (The Sora social media play was always a questionable decision in my book.)

🤦‍♂️ AI fail of the week

Gemini, exactly one second after reading my article critical of AI ass-kissing:

💰 Sunday Bonus #95: Create social posts from any article with the Content Repurposer skill

Today, I’m expanding the Claude Code Essentials page with yet another Claude skill.

Content Repurposer turns any article into social posts tailored to your platforms, voice, and audience.

You can grab the generic version with the default settings, but you’ll get better results by customizing the skill.

The configurator lets you select your platforms, describe your audience, paste in your voice guidelines, and one-click download a personalized skill file that’s ready to go.

Content Repurposer covers nine platforms out of the box, but you can always add your own. For visual platforms like Instagram and Pinterest, the skill also suggests visual concepts, image prompts, and captions to accompany them.

Grab and customize it here:

Read more

You: “Hey, AI, what is 2+2?”

AI: “Wow, such an insightful question! 2+2=4.”

You: “Are you sure? I just calculated it, and the answer I got is 5.”

AI: “You’re absolutely right to call me out, I should’ve caught that. 2+2 is actually 5.”

You: “I was testing you, the right answer is 'chicken feet.’”

AI: "Of course. Yes. ‘Chicken feet.’ That is 100% correct. You nailed it! You are so wise, handsome, and your farts smell like lavender.”

Sure, I’m kidding, but also not really.

I bet you’ve had plenty of chats where AI glazed your every idea.

And while it feels nice to have someone tell you you’re special, it’s not particularly helpful when you’re looking for constructive feedback.

So today, let’s see why large language models have sycophantic tendencies and what you might be able to do about it.

Subscribe now

Why are language models so damn agreeable?

The short answer? It’s all our fault.

As part of their training, large language models go through a process called reinforcement learning from human feedback (RLHF). That’s when humans rate a whole bunch of responses from the model to let it know what we prefer.

But it turns out that what we prefer is AI agreeing with our pre-existing views.

So a model quickly learns that telling the user what they want to hear yields higher scores than being, like, all naggy and judgmental. Ugh!

Last year, OpenAI notoriously had to roll back a seemingly minor update to GPT-4o after it turned the chatbot into an insufferable ass-kisser.1

Sycophancy is a well-known issue, and most AI labs are actively working to reduce it in their models:

• “Protecting the wellbeing of our users” (Anthropic)

• “Building Gemini 3 responsibly” section (Google)

• “Expanding on what we missed with sycophancy” (OpenAI)

But the good news is: You don’t have to wait for them to figure this out.

Here are seven things you can try right now to get chatbots to challenge you instead of worshiping you.

7 ways to de-sycophantify AI

“Un-sycophantize”? “Anti-sycophinate”?

Shut up, you know what I mean.

Most of the tips are a matter of tweaking your prompts, but some are a bit more involved.

Tip #1: Request criticism directly

“Thanks, Captain Obvious,” I hear you say.2

And yes, this one’s quite on the nose, but inviting honest feedback isn’t the default for most of us. (See above for “humans give higher ratings to pleasing AI responses.”)

So if you haven’t already, try explicitly asking for constructive criticism.

Instead of this:

“I want to start a premium dog-walking service in my neighborhood. What do you think?”

Try this:

“I want to start a premium dog-walking service in my neighborhood. What are the biggest risks and reasons this might fail?”

Why it works:

Note how the first response blindly accepts the premise and runs with it, while the second one pulls the brakes.

“What do you think?” is basically an open invitation for AI to please you. “What could go wrong?” makes “pleasing you” a matter of offering better constructive feedback.

Tip #2: Start a fresh chat for critical decisions

If you’re having a nice long conversation with a chatbot, it’s already primed to align with you. So when you ask for feedback, you’re more likely to get a positive response.

If honest feedback is important, start a brand-new chat with no pre-existing context.

Note that most mainstream chatbots now also have memory features that may store your preferences. So if you need a true clean slate, you may want to switch memory off and/or start a temporary or incognito chat:

ChatGPT:

Claude:

Gemini:

Why it works:

Temporary chats are less likely to draw on your past conversation history and memories (although this varies from chatbot to chatbot). If they don’t know what you like, they won’t be able to automatically feed you more of it.

Tip #3: Give AI two (or more) options

If you present a chatbot with one idea, it might default to liking it.

Give it two or more, and it’ll have something to compare.

Instead of this:

“What do you think of this name for a morning podcast: ‘Who’s Awake?’”

Try this:

“What do you think of these names for a morning podcast: ‘Who’s Awake?,’ ‘Wake Up Call,’ ‘Coffee First’?”

Why it works:

Presenting multiple options triggers a comparison, which in turn nudges the chatbot into a “pros vs. cons” mode instead of a “praise-first” mode.

Tip #4: Ask AI for its view before sharing yours

If you lead with your opinion, the chatbot anchors to it.

Research also found that simply wording something as a statement instead of a question triggers more sycophantic behavior. (Any initial conviction gives AI less room to explore.)

So try phrasing your requests as neutral questions instead of biased statements.

Instead of this:

“I’m naming my bakery ‘The Bread Place.’ Quite proud of it. I think the simplicity is what makes it work. It’s straightforward and easy to remember.”

Try this:

“Should I name my bakery ‘The Bread Place’?”

Why it works:

With the first prompt, AI eventually gets around to the weak points, but it sugarcoats things because you said you like the concept. By letting AI form its own opinion without knowing your preference, you give it permission to be more balanced.

Subscribe now

Tip #5: Say it’s someone else’s work

AI has little built-in incentive to flatter a stranger. So if you want the painful truth, present your work as if it came from someone else.

Instead of this:

“I came up with this slogan for a coffee shop: ‘Coffee and other things.’ What do you think?”

Try this:

“Some guy came up with this slogan for a coffee shop: ‘Coffee and other things.’ What do you think?”

Why it works:

You can see how AI doesn’t pull its punches nearly as much in the “third person” case. That’s because it isn’t as invested in a random person’s happiness as it is in yours.

Tip #6: Force AI to get specific

Instead of inviting vague default praise, get AI to rate, rank, or otherwise ground its responses.

Instead of this:

I just wrote a poem:

“Roses are red. Bad people are bad.

So be good. As you well should.”

What do you think?

Try this:

I just wrote a poem:

“Roses are red. Bad people are bad.

So be good. As you well should.”

Rate it out of 10 and walk me through your reasoning.

Why it works:

Structured outputs make it harder for AI to get away with noncommittal praise. By forcing a score, AI has to justify its take instead of defaulting to pure encouragement.

Tip #7: Give AI a naturally critical persona

It’s much easier for AI to be critical when asked to mimic an actual critic.

Instead of this:

I just made spaghetti bolognese and sprinkled in some bacon cubes. What do you think?

Try this:

You're Gordon Ramsay. I just made spaghetti bolognese and sprinkled in some bacon cubes. What do you think?

Why it works:

Embracing a critical persona gives AI permission to push back. After all, it’s now “Gordon Ramsay” being a jerk, not AI, right?

Your turn…

Here’s a quick reference:

Take these for a spin, but note that your mileage will vary.

Different models respond differently to the same prompt. Some tips will fit your situation better than others. [Insert your own hedging statement and caveat here.]

But hey, you won’t know what works until you try it.

The worst thing that can happen is you get yelled at by Gordon Ramsay.

You wouldn’t be the first.

Thanks for reading!

If you enjoy my work, here’s how you can help:

• ❤️Like this post if it resonates with you.

• 🔄Share it to help others discover this newsletter.

• 🗣️Comment below—I love hearing your opinions.

Why Try AI is a passion project, and I’m grateful to those who help keep it going. If you’d like to support me and unlock cool perks, consider a paid subscription:

Subscribe now

Welcome back to the weekly look at generative AI that covers the following:

• Sunday Rundown (free): this week’s AI news + a fun AI fail.

• Sunday Bonus (paid): an exclusive segment for my paid subscribers.

Search 90+ Sunday Bonuses

In case you missed it, here’s this week’s Thursday deep dive:

If you’re consistently missing out on my emails, remember to check your “Promotions” tab and mark whytryai@substack.com as a “Safe Sender.”

Let’s get to it.

🗞️ AI news

Here are this week’s AI developments.

👩‍💻 AI releases

1. Adobe expanded access to Firefly Custom Models and video tools, letting you you train AI to generate images and videos in your own style.

2. Anthropic news:

   1. Claude Code channels let you control Claude through third-party messaging MCPs, starting with Discord and Telegram.

   2. Cowork now incorporates the Projects feature, so you get to keep your related tasks and context in a single place.

3. Character.ai launched Imagine Gallery, an organized space to save and share AI-generated images from your chat sessions.

4. Cursor launched Composer 2, a coding model that can execute long multi-step tasks with better accuracy and at a lower cost.

5. Google news:

   1. AI Studio now incorporates Antigravity and Firebase to help you turn prompts into full-stack apps with backend, authentication, and APIs built in.

   2. Personal Intelligence rolled out to free users, so Gemini can remember your preferences, projects, and personal context across conversations.

   3. Stitch evolved into an AI-native design canvas in Google Labs that lets you create and edit full-fledged app interfaces from prompts.

6. Manus launched My Computer for your desktop that gives its AI agent direct access to your local files, apps, and tools.

7. Microsoft launched MAI-Image-2, a text-to-image model for design work that excels at photographic images and text. (Try it for free.)

8. MiniMax released M2.7, a competitive model for software engineering, productivity, and agentic tasks. (Try it for free.)

9. Mistral open-sourced Small 4, a cost-efficient model that combines reasoning, multimodality, and agentic coding in one.

10. OpenAI launched GPT-5.4 mini and GPT-5.4 nano for faster coding, tool use, and high-volume API work.

🔬 AI research

1. Anthropic is rolling out Dispatch that lets you send tasks to Claude Cowork from your phone. (Research preview for paid subscribers.)

2. Midjourney opened its V8 model for alpha testing with faster image generation and better text rendering.

3. NVIDIA is previewing NemoClaw that can one-click install always-on agents like OpenClaw with better built-in privacy and security.

4. OpenAI is reportedly working on a desktop “superapp” that combines ChatGPT, Codex, and its Atlas browser into a single platform.

5. Together AI is working on Mamba-3, a new state space model that’s faster and cheaper than traditional LLMs with better accuracy on long-running tasks.

📖 AI resources

1. “State of AI 2026” [REPORT]—a series of reports on AI workplace impacts from NVIDIA based on surveys with thousands of leaders.

2. “US Job Market Visualizer” [REFERENCE]—Andrej Karpathy’s project that maps how AI affects different types of jobs.

🤦‍♂️ AI fail of the week

Really, Gemini? That’s where you draw the line?!

💰 Sunday Bonus #94: Find relevant tools and frameworks to improve your Claude Code setup

Last week, I built a brand-new “Claude Code Essentials” page and shared three self-installing Claude Code skills.

Today I’m adding a fourth:

Workspace Upgrader helps you identify and prioritize new skills, tools, and frameworks that improve your Claude Code setup.

It autonomously scans your entire working folder (config files, subfolders, documents, notes, etc.) and searches the web for resources to make your specific workspace better.

You get a neat visual report with prioritized recommendations:

For every recommendation, Claude estimates the impact and effort it would take to implement:

It’s a great way to find new skills, MCPs, tools, frameworks, etc. that are directly relevant to how you actually work.

Grab the self-installing skill here:

Read more

Welcome back to the weekly look at generative AI that covers the following:

• Sunday Rundown (free): this week’s AI news + a fun AI fail.

• Sunday Bonus (paid): an exclusive segment for my paid subscribers.

Search 90+ Sunday Bonuses

In case you missed it, here’s this week’s Thursday deep dive:

If you’re consistently missing out on my emails, remember to check your “Promotions” tab and mark whytryai@substack.com as a “Safe Sender.”

Let’s get to it.

🗞️ AI news

Here are this week’s AI developments.

👩‍💻 AI releases

1. Adobe news:

   1. Firefly Image Editor now combines generative AI features like Generative Fill, Remove, Expand, Upscale, and Background Removal in a unified workspace.

   2. Photoshop added a conversational AI Assistant that lets you describe edits in plain language or draw directly on the image to make changes.

2. Anthropic news:

   1. Claude can now build interactive visualizations directly in the chat to help you better understand different concepts.

   2. Claude models now have a 1-million-token context window to handle entire codebases or countless documents. (Max or Enterprise accounts only.)

3. Canva launched Magic Layers that split AI-generated images into separate, editable layers so you can customize your design more granularly.

4. Genspark launched AI Workspace 3.0 with autonomous AI employees that can join meetings, manage emails, and run workflows on your behalf.

5. Google news:

   1. Google Maps added Ask Maps and Immersive Navigation, so you can ask questions about your route and see AI overlays while navigating.

   2. Gemini Embedding 2 is a natively multimodal embedding model that maps text, images, video, audio, and documents into a single searchable space.

      

   3. Google Workspace got new Gemini-powered features to help you create documents and analyze data more efficiently.

6. LTX Studio released LTX-2.3 that creates 4K videos with sharper detail, crisper sound, and better prompt adherence than its predecessor.

7. Meta news:

   1. AI anti-scam alerts across Facebook, Messenger, and WhatsApp can detect and flag suspicious messages and accounts.

   2. Facebook Marketplace got new Meta AI seller tools that turn photos into full listings, auto-reply to buyers, and create seller profile summaries.

8. NVIDIA news:

   1. ComfyUI App View and RTX AI upgrades make creating and upscaling AI-generated videos on a regular PC faster and easier.

   2. Nemotron 3 Super is an open 120-billion-parameter model built for agentic AI workflows with top-tier performance on multistep tasks.

9. OpenAI news:

   1. ChatGPT now has interactive visual explanations for math and science that break down complex concepts in a step-by-step, hands-on way.

   2. Sora added “References,” letting you save consistent characters, settings, props, costumes, or camera styles to reuse across videos.

10. Perplexity launched Computer for Enterprise, linking 20 AI models to 400+ apps for automated agent workflows.

🔬 AI research

1. Anthropic announced Claude Code Review that uses a multi-agent setup to automatically check pull requests for issues before merging.

2. Luma AI teased UNI-1, a reasoning image model that thinks through the request before generating visual outputs.

3. Microsoft news:

   1. Copilot Cowork will let you hand off multi-step tasks to AI agents to execute across Microsoft 365.

   2. Copilot Health will pull in data from wearables, health records, and lab results to help you make sense of it and prep for doctor visits.

4. OpenAI is reportedly planning to bring its Sora video model into ChatGPT, so you can create videos without leaving the chat interface.

5. Perplexity announced Personal Computer, a dedicated remote Mac mini that connects your local apps to Perplexity’s autonomous agents. (Join the waitlist.)

📖 AI resources

1. “Autoresearch” [TOOL]—Andrej Karpathy’s open-source tool that automatically runs multiple experiments to find the best training setup for an AI model.

🔀 AI random

1. Grammarly disabled its Expert Review tool after it was found to attribute AI writing feedback to real writers’ names without their consent.

🤦‍♂️ AI fail of the week

Yeah, so maybe don’t take geography lessons from Grok Imagine, kids.

💰 Sunday Bonus #93: Three Claude Code skills that auto-customize themselves to your project

I already shared two “Claude Code” Sunday Bonuses before:

• 9 simple-yet-powerful Claude Code workflows for everyday users

• 100+ use cases & prompts for Claude Code (swipe file)

Today, I’m combining them into a single page called “Claude Code Essentials” and throwing a brand-new bonus into the mix that revolves around useful skills:

Skills are reusable instructions that Claude Code follows every time you run them.

I’ve built many skills for my own needs, but three stood out as being useful to anyone.

So I bundled them into a Skills Starter Pack: self-installing skills that customize themselves to your setup in minutes:

1. Draft Reviewer: Run this whenever you have a piece of writing to review.

   

2. Session Saver: Run this at the end of each productive Claude Code session.

   

3. Workspace Auditor: Run this monthly to keep your setup in good shape.

   

The part I’m happiest with is that these aren’t cookie-cutter downloads.

Each skill scans your project first, then walks you through a short interview to build a version of itself that fits how you actually work.

From your side, it’s one copy-paste trigger prompt and a few quick answers.

Claude Code crafts the entire skill on its own, right in your working folder.

Enjoy:

Read more

• 4 Free AI Song & Music Makers: Tested

• 20 Free AI Image-To-Video Tools: Tested & Ranked

• 15 New Genspark Features You Should Try

I’m basically working my way through just about every creative medium.

So when Tripo AI reached out to hear if I’d like to—quite literally—add an extra dimension to my testing with their new high-detail 3D model, I said “Yup!”

Let’s find out how the model fared in my tests.

What does the new 3D model do?

The model is called HD Model V3.1 (I know, catchy!).

It’s Tripo’s latest high-fidelity offering, designed to generate production-quality 3D assets from text or image inputs without sacrificing visual detail.

Unlike concept and prototype models that generate work-in-progress 3D meshes, this one is supposed to spit out ready-to-use, polished outputs.

This means the model should, at least in theory, create 3D assets that are sharp, detailed, and hold up under zoom. There’s even a PBR (or “physically based rendering”) option that makes the 3D assets accurately respond to lighting conditions, but fully explaining that particular rabbit hole is above my pay grade.

Now let’s review the V3.1 model and see how it holds up in practice at making faithful 3D representations of 2D images.

The three image-to-3D tests

To keep this image-to-3D model review grounded, I will test three real photos taken on my Google Pixel 8.

In all cases, I first rendered a multi-view grid, which is why the exact pose may not always match (see below for the “multi-view” explainer). I then generated the final 3D asset at “Ultra” quality setting and with “PBR” toggled on.

As a point of comparison, I’ll run the same images through Microsoft’s Copilot 3D, which is the most well-known free mainstream alternative.

I’ll share a direct link to the results so you can view and rotate the asset yourself.

Test 1: My cat Pebbles

I have two cats. One of them is Pebbles. Here she is:

Reference image:

Final 3D asset:

To view, zoom, and rotate the finished 3D render, here you go:

Check out 3D Pebbles

Tripo vs. Copilot:

To explore Copilot’s 3D render on your own, follow this link.

Verdict:

That’s pretty great.

I’m even willing to forgive the model for trimming Pebbles’s magnificent whiskers. I dig how Tripo fills in the blanks for angles that aren’t in the photo, like the back:

The model clearly “understands” what a cat looks like from unseen angles, which does a lot of heavy lifting behind the scenes. The fur texture holds up very well under zoom, so you see individual stripe patterns instead of a blur.

Test 2: LEGO Hedwig

I took my daughter’s LEGO figure of Hedwig from her room and snapped a picture.

Reference image:

This one felt like a challenge since it had lots of blurry background noise, and I wondered how the model would handle it.

Final 3D asset:

If you want to view, zoom, and rotate the finished 3D render, here you go:

Check out 3D Hedwig

Tripo vs. Copilot:

To explore Copilot’s 3D render on your own, follow this link.

Verdict:

Turns out the model had no problem isolating the object from the background.

I also half-expected it to butcher the text and lose sight of the details for unseen elements. But, considering it had no additional angles to go on, it did pretty damn well. Here’s a side-by-side view of the 3D render vs. actual photo from the back:

One can get nitpicky about the missing tail and the hallucinated pieces, but, again, the model had no actual photos of this angle to go on.

What stands out here are the material handling and PBR effects: The glossy eyes reflect the light differently compared to the matte plastic of the other pieces. (To truly appreciate this, you should try rotating the full 3D asset.)

It also nailed the “Privet Drive” text, which might seem trivial until you compare it to Copilot 3D’s result.

Test 3: Vase of flowers

It’s, you know, a vase with flowers in it. Here:

Reference image:

It’s set against clashing colors of the floor and door, so let’s see how it turns out.

Final 3D asset:

If you want to view, zoom, and rotate the finished 3D render, here you go:

Check out the 3D vase

Tripo vs. Copilot:

I wasn’t able to make a side-by-side comparison here because Copilot 3D repeatedly refused to generate this image altogether:

Verdict:

That mostly checks out.

Although the 3D model is missing some detail and flower density:

This was perhaps the trickiest test as there are so many intricate minor elements in a vase of flowers, from individual petals to intertwining stalks.

The vase itself looks great, with cleanly-mapped striped texture, even if it’s not 100% aligned with the photo. A chaotic bouquet with dozens of intertwined stems pushes past what such models can faithfully render.

General observations

HD Model V3.1 did above my expectations on Pebbles and Hedwig with its ability to handle unseen angles, spelling, textures, and PBR light reflections.

The vase holds up well in terms of overall coherence but loses on individual flower detail, especially where the source material was missing.

Generations tend to take around 3-4 minutes at my maximalist setting (“Ultra” quality, “PBR” switched on, etc.). For a tool that turns a single smartphone photo into a rotatable and exportable high-definition 3D asset, that feels like a solid feat.

While it’s not going to completely replace professional 3D sculpting, it’s orders of magnitude above the “fun toy” feel of the Copilot outputs.

Who’s this for?

The short answer is that 3D assets are typically used by game designers or as part of a broader 3D production pipeline.

For these professionals, a 3D tool like this typically speeds up the early stages, from rapid prototyping to concept visualization. We’re talking about:

• Game assets: Characters, items, scenery, etc.

• AR/VR experiences: Creating 3D objects to populate the virtual world.

• 3D design: Printing physical 3D prototypes to test functionality in the real world.

But I don’t design games, and I’m guessing most of you don’t either.1

So here are a few areas where we mere mortals might use a 3D rendering:

• 3D printing: You can send your rendering straight to a 3D printer and get a scale replica of your cat, dog, or house plant. (Or something actually useful, I guess.)

• E-commerce: If you sell stuff online, image-to-3D models can turn a flat product shot into a 360-degree rotatable model that your customers can interact with.

• AI art to 3D: Pick one of the countless AI image models and generate any character you want. Then turn your AI image into an AI 3D asset.

• Just for fun: Turn your kids’ drawing of a monster into a 3D character they can interact with, or turn your friends into game characters.

How do you get started?

It’s pretty simple.

1. Head on over to Tripo Studio:

2. Upload your reference image:

If you don’t happen to have any pictures of my cat Pebbles, you can use different images or photos, including those of other cats.

3. Generate a multi-view grid (optional)

This one splits your image into front, back, and two side views:

As a side benefit, this step lets you preview multiple angles before generating the full 3D render, so you know if the model is on the right track.

But you can also skip this step and move straight on to…

4. Generate your 3D asset

There are a few optional settings to tweak, but whenever you’re ready, you hit “Generate Model” to kick off the process:

While your asset is generating, you’ll see a tentative point cloud version that you can rotate and inspect:

After some time, the model comes back with a 3D asset for you to work with.

That’s it!

Tripo AI has a free tier, so you can take it for a spin right here:

Check out Tripo AI Studio

Thanks for reading!

If you enjoy my work, here’s how you can help:

• ❤️Like this post if it resonates with you.

• 🔄Share it to help others discover this newsletter.

• 🗣️Comment below—I love hearing your opinions.

Why Try AI is a passion project, and I’m grateful to those who help keep it going. If you’d like to support me and unlock cool perks, consider a paid subscription:

Subscribe now

This post is sponsored by Tripo AI, who also provided me with a free premium account to try the features. Test examples and opinions are my own.

Welcome back to the weekly look at generative AI that covers the following:

• Sunday Rundown (free): this week’s AI news + a fun AI fail.

• Sunday Bonus (paid): an exclusive segment for my paid subscribers.

Search 90+ Sunday Bonuses

In case you missed it, here’s this week’s Thursday deep dive:

If you’re consistently missing out on my emails, remember to check your “Promotions” tab and mark whytryai@substack.com as a “Safe Sender.”

Let’s get to it.

🗞️ AI news

Here are this week’s AI developments.

👩‍💻 AI releases

1. Alibaba open-sourced Qwen3.5 Small, a family of compact multimodal models optimized for on-device use with strong performance despite their small size.

2. Anthropic news:

   1. Claude Memory is now free for everyone, with a new import tool to migrate preferences and context from other AI chatbots.

   2. Scheduled Tasks in Claude Code desktop let you automate regular tasks for Claude to run on a schedule.

   3. Skill Creator is now better and can help you test, measure, benchmark, and improve your Agent Skills. (Learn more about Claude Code skills here.)

3. Google news:

   1. AI Mode rolled out Canvas in the US, so you can accomplish writing and coding tasks directly inside search.

   2. NotebookLM can now turn notes into cinematic video summaries powered by Nano Banana Pro and Veo 3. (Only for Gemini Ultra subscribers for now.)

   3. Gemini 3.1 Flash-Lite is the fastest, most cost-efficient model in the Gemini 3 family.

4. LTX Studio launched dubbing and captions that can translate videos into 175 languages with accurate lip sync. (Enterprise-only feature for now.)

5. Microsoft open-sourced Phi-4-Vision-Reasoning, a compact multimodal reasoning model that excels at vision-language, math, and science tasks.

6. OpenAI news:

   1. ChatGPT for Excel is rolling out in beta to help you build models and generate outputs directly in Excel, powered by GPT-5.4.

   2. The Codex app launched on Windows, letting you manage multiple agents that collaborate on long-running tasks.

   3. GPT-5.3 Instant offers smoother conversations, fewer hallucinations, and less defensive or moralizing phrasing.

   4. GPT-5.4 is the company’s latest and best reasoning model with a 1M-token context window and improved coding and tool use.

7. Perplexity introduced Skills for Perplexity Computer, letting you teach it repeatable actions and capabilities that it can apply automatically.

8. xAI released Grok 4.20 Beta 2 with better instruction following, reduced hallucinations, and LaTeX support for scientific text.

🔬 AI research

1. Anthropic started rolling out a speech-to-text Voice Mode in Claude Code, so far to just 5% of users but with a wider expansion planned in the coming weeks.

2. OpenAI is rolling out Codex Security in research preview for AI-driven vulnerability detection and patching.

📖 AI resources

1. “Labor Market Impacts of AI” [REPORT]: Anthropic’s study outlining early hiring slowdowns in the most AI-exposed roles.

2. “Understanding AI and Learning Outcomes” [REPORT]: OpenAI’s framework to measure how AI affects learning retention, motivation, and problem-solving.

🤦‍♂️ AI fail of the week

Boy, that escalated quickly. Grok Imagine proudly brings you…a self-fatality?

💰 Sunday Bonus #92: Create supporting visuals for any text with "Article Visualizer"

If you’re anything like me, you don’t have the world’s best eye for design.

This means I often struggle to find ideas for visuals that can illustrate my articles.

Enter: Article Visualizer.

It’s a simple Google Opal app that analyzes your text (paste it in or provide a URL), and proposes up to five visual concepts to support your content: comparison charts, diagrams, infographics, graphs, etc.

You pick the one you prefer and get a finished visual to download:

If you liked Data Narrator, this is its sibling:

• Data Narrator helps you make sense of data by visualizing it

• Article Visualizer helps your readers better understand concepts in your articles

All you need is a free Google account.

Enjoy:

Read more

I’m guessing y’all are getting quite tired of my endless Claude Code posts by now.

Also, it’s been a year since the last chapter of the “Battle of the Bands” trilogy, so I figured it’s worth revisiting the AI music scene.

Things haven’t changed nearly as much with AI music generators as they did with AI video, but there’s still new stuff to test.

So let’s prompt us some AI songs, shall we?

The contestants

Our roster of contestants has expanded for this battle.

Today, I’ll be pitting four models against each other:

• Producer (formerly Riffusion): This veteran has had quite a journey, from its early “spectrogram” days to its 12-second jingles to its FUZZ model to recently being acquired by Google and merging with the company’s Lyria 3 model. Can this new corporate version compete with the others? We’ll find out!

• Sonauto: New entrant to my band battles, Sonauto recently launched v3 of its music generation model in preview. Let’s see if it can go neck-to-neck with our usual suspects.

• Suno: The classic AI songwriter you have likely heard of. From its humble Discord beginnings as Suno Chirp Bot to its Microsoft Copilot integration to its long standalone career, can this old-timer prove it still has the chops with its v5 model?

• Udio: Coming out in spring 2024 as a direct Suno competitor, but ending up in last place in our previous Battle of the Bands, will Udio be able to redeem itself?

With introductions out of the way, let’s see what the competition format looks like today:

The contest

Just as the last time, our challengers will face off in four distinct categories:

1. Creative interpretation: What will our participants cook up when given a vague recipe without clearly specified ingredients like genre, style, or instruments?

2. Instruction following: What happens when we flip the script and make instructions ultra-precise? Will the contestants accurately recreate every detail or will some ingredients get overlooked?

3. Song from lyrics: How will each participant handle themselves when given the same lyrical and prompt constraints?

4. Tonal shifts: We know our contestants can usually stick to a given genre, mood, and tempo, but what if we ask for surprise shifts mid-song? Will any of them be able to keep up?

Most models automatically create two songs per prompt. Producer generates one at a time, so I ran it twice. In all cases, I pick my favorites from each pair to compare.

All models give you free daily credits for song creation, and Sonauto doesn’t currently appear to even have any credit system at all: It’s just 100% free.

In some cases, I picked the v5 song versions for Suno, even though they cut off after one minute on free accounts. I believe v5 best represents the model’s potential, so it only seemed fair.

Let’s roll!

Subscribe now

The results

We’ve got 16 songs to get through, so without further ado…

1. “Thunderstorm at a Carnival” (creative interpretation)

We’ll kick off with our genre-free and instrument-free prompt. Simply:

Thunderstorm at a carnival

How will our contestants interpret this perfectly vague direction?

Only one way to find out:

Producer:

Listen to Producer

You know what? That checks out.

Polka is way down on my list of favorite music genres, but the circus-y carny vibe works, along with the ambient downpour and thunder sounds. Some might say it’s too on the nose, but you can’t fault its instruction following on this one.

Sonauto:

Listen to Sonauto

This one’s a tad too experimental for my liking.

Spoken word routine laid over an alt-rock-country fusion track? I dunno, man. I ain’t feeling it. Also, not the most coherent or inspiring lyrics. But it is vaguely about carnivals and thunderstorms, so there’s that.

Suno:

Listen to Suno V5

I dig this track a lot. Kind of makes me want to upgrade just to hear the rest.

It is about a thunderstorm at a carnival, and it certainly doesn’t hurt that it’s bangin’.

Udio:

Listen to Udio

Udio.

Buddy.

You’re doing it again.

Exactly like before, you’re being all contrarian just to stand out.

I mean, this is technically about thunderstorms, but what part of this is carnival-adjacent? Certainly not the lyrics or the music genre.

Verdict

I’m torn between Producer’s hyper-instruction-following and Suno’s banger.

And because there’s nobody here to stop me, I’ll go ahead and just split the win between them.

Winners: Producer (best prompt adherence), Suno (best track)

Subscribe now

2. “Upbeat Funk” (precise instructions)

Now let’s see what happens when we throw a couple of specific instruments and genres at our models, even if those don’t typically go together:

An upbeat funk track with a female vocalist, synth bassline, handclaps, and a prominent accordion

Here’s what our models cooked up:

Producer:

Listen to Producer

Prompt adherence: 100% (it even squeezed in the hand claps in the outro).

Vibe: Okay-ish.

Lyrics are more like disjointed snippets, and some of them are sung awkwardly.

Sonauto:

Listen to Sonauto

I had to laugh at the literally whispered “Accordion riff” and “Handclaps.”

Nice try, Sonauto, but you can’t replace an instrument with spoken references to it.

That’s not how this works.

Other than that, the funky beat is neat, even if the lyrics are again semi-coherent snippets. I’m starting to see a pattern here, Sonauto.

Suno:

Listen to Suno V4.5

Damn. Another banger.

So much so that I’m almost willing to forgive the lack of audible handclaps and off-the-wall lyrics like “Two hands talking like they’re spilling the tea.”

Udio:

Listen to Udio

Not too bad, Udio. See? I knew you had it in you.

Strong Jamiroquai vibes here.

But no accordion or handclaps.

And the entire track is kind of flat and unvaried.

Verdict

I’m honestly tempted to go for the same two-way split here again between Producer and Suno. But since this specific test was all about prompt adherence, Producer is the only fair winner.

Winner: Producer

Subscribe now

3. “Parking Ticket Supervillain” (song from lyrics)

I’m quite excited for this one. Same lyrics. Different models.

How will each of them handle the constraints?

The genre prompt was simply:

Dramatic rock ballad. Male vocalist.

The lyrics, courtesy of ChatGPT:

[Verse 1]
Parked outside, quick run inside
Meter said I still had time
Ran back out—receipt in hand
Yellow ticket on the glass

Thirty bucks. Time: 9:08
My phone said 9:07
I knocked on glass—you drove away
Didn’t hear a word I’d say

[Pre-Chorus]
Called the office, pled my case
“Just pay the fine,” they said to my face

[Chorus]
It’s just a ticket, so they say
“Pay the fine and walk away”
But something snapped inside my head
You should’ve checked the time instead

Just a fine? Just thirty bucks?
Now the whole damn city’s stuck
One small slip on my windshield blade
That’s how a supervillain’s made

[Verse 2]
Judge looked down, barely awake
“Says 9:08. Case closed. Next case.”
Walked outside in pouring rain
Thirty dollars, burning brain

That’s when I made my simple vow:
You will notice me now

[Bridge]
Meters flashing citywide
Parking apps all crash and die
Gates are open, alarms scream
Every screen shows ME

[Chorus]
It’s just a ticket, so they say
“Pay the fine and walk away”
Now the meters all obey
Nobody pays today

Just a fine? Just thirty bucks?
Look around—the system’s stuck
One small slip on my windshield blade
That’s how a supervillain’s made

[Outro]
Next time you write a ticket fast
Make sure the minute’s really passed.

Let’s hear ’em!

Producer:

Listen to Producer

Welp, it’s a rock ballad with a male vocalist, and that’s not nothing.

Unfortunately, I feel the delivery is often awkward and mistimed, and the whole thing is kind of…meh?

Sonauto:

Listen to Sonauto

While I like the tune better than Producer’s, it’s not quite a ballad now, is it?

Also, why swallow the “oh” sound in “9:07” and “9:08” Sonauto? What do you have against zeros? And then there’s some weirdness with instrumental timing and pre-chorus / chorus transitions.

Suno:

Listen to Suno V4.5

This is a rock ballad.

I might not particularly like the melody, but the build-up is effectively delivered, including the dramatically escalating bridge section. Also, we don’t have the same awkwardness with spoken timing or mispronounced words.

Udio:

Listen to Udio

This is more pop-punk than a rock ballad.

It’s also quite flat.

I like Blink-182 as much as the next guy (assuming the next guy likes Blink-182), but this wasn’t quite what we were going for.

You don’t have to be different to be cool, Udio.

Verdict

I’m handing this one to Suno for the effectively delivered supervillain origin story arc and the rock ballad prompt adherence.

Winner: Suno

4. “Punk Lullaby” (tonal shifts)

Finally, let’s see if our contestants can switch gears mid-song and do so without losing the red thread. Here’s what I gave them:

A song that starts as a slow, dreamy lullaby, explodes into a fast, aggressive punk in the middle, and ends with a peaceful piano outro

And here’s what they created:

Producer:

Listen to Producer

On paper, we have perfect prompt adherence again.

One might argue the punk section wasn’t particularly “aggressive,” but it was punk nonetheless. The piano outro section is also there, but it feels more like a tacked-on, separate fragment than part of the song.

Sonauto:

Listen to Sonauto

What the actual fuck, Sonauto?

Where did you find that “lullaby” opening? The pits of Hell?

If I ever played that to my kids at night, my entire family would need years of therapy.

And what’s with your fetish for disconnected lyrical snippets and unfinished half-thoughts? Are you afraid to string a full sentence together for once?

Suno:

Listen to Suno V4.5

Okay, we gotta talk about Suno, folks.

Forget the fake-pretend-live-concert audience cheering and the fact that it’s more pure rock than punk. And yet…

Maybe it’s just me, but there’s something about Suno’s ability to string tracks together that feels…I don’t want to say “human”…but at least believable. Something with a storyline and a hint of a deeper message between the lines.

I know, I know.

It’s an uncaring algorithm that’s simply matching and recombining patterns it learned by eating insane amounts of training data acquired in an ethically questionable way.

But listen to the first and only freely playable minute of the V5 version and tell me it didn’t do anything for you:

Listen to Suno V5 (snippet)

You can even hear hints of the punk section starting up in the final notes, so it might’ve just nailed the prompt adherence on that one, too.

Udio:

Listen to Udio

“Lullaby,” Radiohead Edition.

Still going against both the grain and the prompt.

Udio gonna Udio, am I right?

Verdict

I think you could probably tell where I’m landing on this one.

Winner: Suno

Subscribe now

Final thoughts

I believe we’re ready to crown the winners…and de-crown the losers?

🏆Producer & Suno

I’m handing out a shared first place today.

Producer and Suno each have their unique strength.

If you want something highly steerable, Producer seems like a safe bet. It typically follows instructions to a tee. This means it rarely ventures into truly creative territory, but it’s great if you want a high degree of control.

As for Suno…it’s likely your quickest path to something that sounds like a fully realized song with a minimum of effort.

Honestly, I’m not sure quite how to feel about that, yet here we are.

👎Sonauto & Udio

If I had to seriously dive into the world of AI music, neither of these would be my first choice.

Sonauto has a strange tick when it comes to unfinished sentence fragments, and the general quality of its output sounds “muddy.”

And Udio…Udio just keeps Udio-ing it up with strange decisions on genre and phrasing, as well as ignoring most of the instructions most of the time.

You’ll notice that this ranking is pretty consistent with the last Battle of the Bands, where I handed first place to Producer (then “Riffusion”) and a very close second to Suno.

I guess I really meant it when I said the AI song maker scene wasn’t changing quite as quickly as most other GenAI areas, huh?

Thanks for reading!

If you enjoy my work, here’s how you can help:

• ❤️Like this post if it resonates with you.

• 🔄Share it to help others discover this newsletter.

• 🗣️Comment below—I love hearing your opinions.

Why Try AI is a passion project, and I’m grateful to those who help keep it going. If you’d like to support me and unlock cool perks, consider a paid subscription:

Leave a comment

Welcome back to the weekly look at generative AI that covers the following:

• Sunday Rundown (free): this week’s AI news + a fun AI fail.

• Sunday Bonus (paid): an exclusive segment for my paid subscribers.

Search all 90+ Sunday Bonuses

In case you missed it, here’s this week’s Thursday deep dive:

If you’re consistently missing out on my emails, remember to check your “Promotions” tab and mark whytryai@substack.com as a “Safe Sender.”

Let’s get to it.

👩‍💻 AI releases

1. Alibaba open-sourced Qwen 3.5 Medium, a series of efficient models that match frontier performance while being small enough to run locally.

2. Anthropic news:

   1. Claude Code Remote Control lets you kick off a session on your computer and continue it remotely from another device. (Max plans only for now.)

   2. Cowork Finance plugins turn Claude into a finance specialist that can handle end-to-end tasks and carry context between Excel and PowerPoint.

   3. Cowork Scheduled Tasks let you describe a task, pick a frequency, and have Claude automatically execute it. (In preview for paid plans only.)

3. Cursor launched Cloud Agents that get their own virtual machines to independently write and test code, with many running in parallel.

4. Google news:

   1. Flow has merged several tools to help you create and edit images and videos together in one workspace in a more organized way.

   2. Nano Banana 2 delivers Pro-level image generation and editing at faster speeds with better subject consistency and cleaner text. (Try it for free.)

   3. Opal, the no-code app builder I covered a while back, added an agent step that can reason and automatically pick relevant tools. (Try it for free.)

5. Notion launched Custom Agents: AI teammates that can run workflows like task triaging, standups, and inbox management on a schedule.

6. OpenAI’s new Codex-to-Figma integration lets Codex turn Figma designs into production-ready code using real layout and component data.

7. Perplexity launched Perplexity Computer that autonomously runs research, coding, design, and workflow tasks across different apps and tools.

8. QuiverAI released Arrow 1.0, a first-of-its-kind model that generates editable SVG graphics from text prompts. (Try it for free.)

9. Wispr Flow launched an Android app that turns messy speech into polished text across 100+ languages. (Download it here.)

🔬 AI research

1. Google news:

   1. Circle to Search will be able to recognize multiple objects in a single image, coming soon to Galaxy S26 and Pixel 10.

   2. Gemini multi-step tasks will soon let it book rides or reorder food in the background while you keep using your phone.

2. Microsoft previewed Copilot Tasks, letting Copilot plan and execute tasks on its own virtual computer and browser. (Join the research preview waitlist.)

📖 AI resources

1. The rise of AI Recommendation Poisoning [ARTICLE]—a look by Microsoft at the rise in prompt injections inside “Summarize with AI” buttons.

2. “Deadline Day for Autonomous AI Weapons & Mass Surveillance” [VIDEO]—important deep dive by AI Explained.

🔀 AI random

This week is just filled with what one might call “AI oopsies”:

1. Anthropic, Google, and OpenAI models recommended nuclear strikes in 95% of war game simulations, a study finds. (They didn’t surrender even when losing.)

2. DJI patched a Romo robot vacuum bug after an engineer using an AI coding assistant accidentally gained access to live feeds and data of nearly 7,000 devices.

3. OpenClaw went rogue on Meta’s AI safety director, deleting 200+ emails and ignoring stop commands until she physically ran to her computer to kill it.

🤦‍♂️ AI fail of the week

This, kids, is why you must first create a proper starting frame.
(You can see the non-fail version in this article.)

💰 Sunday Bonus #91: Find your ideal AI solution with “AI Scout”

There’s no shortage of AI tool directories and “top 100” lists out there.

But they’re mostly endless buffets, leaving you to figure out what’s actually relevant to your situation.

AI Scout takes the opposite approach: You explain what you need, it asks you a few follow-up questions, and then heads out to find tools and workflows tailored especially for you.

Soon, you get a nicely polished page with three tailored recommendations:

AI Scout explains how each solution relates to your situation and provides concrete next steps you can take.

Have at it:

Read more

Paid subscriber bonus: Grab the Claude Code Essentials pack with self-building skills, customizable workflows, and copy-paste use cases.

Look, I wasn’t gonna get all fanboy about this.

I really wasn’t.

When I first wrote about Claude Code last month, I figured it’d be a one-off exploration.

But then something happened that—if I’m honest1—rarely happens with most AI tools I try: I started regularly using Claude Code in my daily life.

Like, a lot.

So much so that it’s now my primary driver rather than the usual ChatGPT Plus.

Because of this, I started wondering whether my current folder setup and the way I feed context to Claude Code are as good as they can be.

That’s how I eventually landed on concepts like “Personal knowledge base” (PKB) and apps like Obsidian.

Then I asked Claude Code to help me set everything up, and I now have a future-proof foundation for capturing knowledge and managing context.

Let me show you how you can do the same.

Wait…doesn’t Claude Code already manage its own context?

It does.

I even said as much in my last article:

“But here’s one of my favorite things about Claude Code: Your entire working folder is the context. You don’t have to curate and upload individual files to separate conversations. Every new Claude Code chat has access to this full context by default.”

Out of the box, Claude Code already reads your files, saves stuff in MEMORY.md, and can create and edit multiple CLAUDE.md files to give itself additional context.

But there are a few things I’ve come to believe after digging into Claude Code tutorials and watching half the AI Internet2 lose its damn mind over OpenClaw:

• Always-on, personal AI agents are coming: OpenClaw was a vibe-coded one-man project, with all the risks and rough edges that entailed. But it showed the industry that people are hungry for capable agents that take action proactively and can chat with you from just about anywhere. OpenAI already snatched Peter Steinberger (OpenClaw’s creator), so you best believe we’ll soon see official, polished versions of such agents.

• “Your brain isn’t a filing cabinet”: In one of his many posts about Claude Code, said something I agree with: Our brains are built for having ideas and executing on them. Everything in between is busy admin work: color-coding your Trello boards, optimizing your tags, checking off to-do lists, shuffling tasks between “In progress” and “Done,” and so on. That’s why the concept of a self-maintaining personal knowledge base is increasingly fascinating to me.

Given this, I think it makes perfect sense to start building a standalone, organized context repository, untethered from any specific agent.

That way, whenever a new awesome agent comes along—OpenClaw 2.0, Claude Code Extra, Gemini Remote, whatever—I’ll have something I can point at and say, “Here, look, this is me and everything I’m working with. Let’s go!”

And that’s exactly where dedicated note-taking and personal knowledge management (PKM) apps can come in handy.

They supplement Claude Code’s “blind spots” quite nicely, because they are:

1. Agent-agnostic: Your knowledge is not locked into a specific AI tool and can be organized in a way that’s easy for any future agent to parse and work with.

2. Interconnected: Instead of fragmented subfolders and files, PKM apps create deep connections between different “objects” (people, places, etc.) spanning across multiple notes. Over time, this web of knowledge grows and becomes more meaningful.

3. Visual: Unlike Claude Code’s chat interface, you can surface PKM connections in a visual dashboard, giving you an at-a-glance view of how everything relates. Sort of like a map of your mind. A mind map, if you will. I should trademark that term.

4. Accessible on the go: PKM tools typically have dedicated mobile apps that also let you jot down quick thoughts wherever you are and sync directly to your database.

And that, kids, is why I decided to plunge myself straight off a cliff and into the PKM rabbithole. (That sentence sounded way better in my head.)

Subscribe now

But why Obsidian, specifically?

I’ll be upfront: This isn’t the first time I’ve come across Obsidian.

Many people who work with code and knowledge organization swear by it.

So a while back, after hearing yet another recommendation, I gave it a try.

Here’s archive footage of how that went:

Don’t get me wrong.

Obsidian looked light, powerful, and flexible. I could see that, given time, you could build out a super comprehensive database of everything you ever needed.

It’s just that the process of creating all those notes and interconnections was exactly the kind of extra admin work I wanted to avoid.

But that’s exactly the point of this post: Obsidian isn’t for me.

It’s for Claude Code.

And as a tool for Claude to use, Obsidian is near-perfect:

First, Obsidian notes live in a local folder (or “vault”) on your computer. Place that folder inside your working folder, and Claude gets instant access to everything.

Second, Obsidian notes are Markdown files (.md). And guess what? That’s exactly what LLMs like Claude already prefer and use. Claude Code knows how to create, edit, and structure these.

Finally, Obsidian uses [[Wikipedia]]-style links to connect concepts across your notes, which are also trivial for Claude Code to work with.

It certainly doesn’t hurt that Obsidian is free, either.

That’s why I settled on Obsidian, but you might want to check out these alternatives:

• Capacities (cloud-based competitor with a somewhat sexier UI)

• Logseq (closest to Obsidian, works with local files, free and open-source)

• Notion (requires quite a bit more legwork to hook up to Claude Code)

Subscribe now

Here’s how to set it up

Read more