Anthropic’s new top-tier model (Claude Fable 5/Mythos 5) is way too pricey and comes with access restrictions, but most of us don’t need it anyway.

What is it?

Fable 5/Mythos 5 is Anthropic’s latest model and a step change for frontier LLMs.

Before we move on, let’s clear up the whole Fable vs. Mythos thing, so I don’t have to keep writing them in tandem:

• Fable 5: Is the same underlying model as Mythos 5 but with additional safeguards in place that automatically reroute certain sensitive queries (e.g. cybersecurity and biology) to Opus 4.8 instead. Available to everyone.

• Mythos 5: The “full” version without any safeguards. Available only to a vetted group of “cyberdefenders and infrastructure providers.”

Claude Fable 5 leads the pack on virtually every benchmark self-reported by Anthropic:

But you don’t have to take Anthropic’s word for it.

The emerging consensus is that Fable 5 truly is a leap forward across the board.

Opinions aren’t nearly as divided as they’ve been for Anthropic’s latest releases, and even the notoriously critical Reddit crowd seems broadly positive.

You can find real-world tests and examples below in “Further reading & watching,” but the short of it is: Fable 5 is a state-of-the-art model that excels at the hardest, most complex long-horizon tasks and agentic coding work.

Subscribe now

How do you use it?

Do you have a paid Claude subscription?

Congratulations: You already have access to Claude Fable 5.

Just open your Claude app or claude.ai, and you’ll see Fable 5 in the model picker:1

Go crazy!

Ah, but you must hurry, fellow explorer: Regular Claude subscribers only have until June 22 to test this mythical model. For from June 23 onward, Fable 5 shall only reveal its powers to those willing to pay its legendary per-usage fees.2

Even before the June 23 switcheroo, Fable 5 will burn through your daily/weekly rate limits at twice the speed of the already expensive Opus 4.X model family.

To summarize, Fable 5 is:

• “Mythos 5 Lite” with safety guardrails that limit its range of use cases.

• Only available in existing paid Claude plans until June 22.

• Anthropic’s most expensive model (double the cost of Opus 4.8).

Looks like the “AI rich vs. AI poor” future predicted is finally here.

But wait…how much does this really matter to you?

Subscribe now

Why should you (not) care?

Okay, real talk.3

For the overwhelming majority of us, Fable 5 is pure overkill.

Unless you do frontier research or work on complex software engineering projects like Stripe’s reported codebase migration, you simply don’t need Fable 5’s wizard skills.

Take a look at these ELO scores for GDPval-AA, which “evaluates AI models on real-world, economically valuable tasks across a wide range of occupations”:

Now be honest: Do Fable’s extra 42 ELO points over Opus 4.8 really matter for your daily tasks? How would you even measure it? Are those marginal gains worth the 2x price hike?

I’ll take a wild swing here and say: No, they aren’t. Not for the average person.

For instance, I consider myself to be at least an intermediate AI user. I work with and test AI tools daily to keep up with all the madness for this newsletter.

I run Claude Code together with Codex, hooked up to my Obsidian vault as their command center.

I’m in the weeds with this stuff, is what I’m saying.

But for most of my use cases, I find that the good old Opus 4.64 is more than plenty. Hell, even Sonnet 4.6 works well for much of what I need done.

(I do ground-breaking scientific research only occasionally and have to date cured at most two or three cancers.)

And it’s not just me, either.

In their recent Fable 5 “vibe check,” Dan Shipper and the crew at Every wrote:

…we found that users who were highly adept with AI—at Level 7 or 8 on our AI adoption ladder—found [Fable 5] paradigm-shifting for their hardest tasks. Users who were lower down on the curve, however, struggled to find something to use it for.

For reference, here are the 8 levels they’re talking about:

So if you’re not “managing multiple long-running agents at the same time,” it’s perfectly fine to give Fable 5 a pass.

You don’t need Fable 5 to draft a quick work email or suggest entertainment ideas for your pet iguana’s birthday party.

Fable 5 is also way overqualified for most of your everyday knowledge work.

It’s like grabbing a bazooka to take down a mosquito.

It’s like taking a Formula 1 car to drive your kids to school.

It’s like asking Fable 5 to come up with a third beat for this three-part analogy.

There’s also early anecdotal evidence that on certain white-collar tasks like writing and front-end design, Fable 5 fares worse than cheaper models due to its dense engineer-style lingo and verbosity.

“But Daniel, I saw that Fable 5 can flawlessly code entire games and build advanced tools from a single prompt! Isn’t that worth something?” you ask, conveniently setting up my upcoming response.

And yes, that appears to be a legitimately magical use of this tech. But:

a) Let’s face it: Most of us aren’t sitting on a backlog of brilliant ideas we’re itching to will into existence. (Although now that the option is available, this may gradually change, which is a net positive for all of us.)

b) One-shotting complex software is bound to push your token spend into the stratosphere, which brings us full circle to square one.5

As it stands, Fable 5 is likely not for me and you…and that’s okay.

Take Fable 5 for a spin on your regular Claude plan until June 22. Poke at it, see what it can do and how it feels compared to what you’re used to.

Then ask yourself if you’re willing to pay double the price—and wait longer for Fable to painstakingly think through every request—instead of sticking to Sonnet or Opus.

My guess?

We won’t be too heartbroken when it’s time to say goodbye to Fable 5 in two weeks.

Further reading & watching

• “Anthropic Just Split the Frontier in Two“ - Product with Attitude

• “Claude Fable 5 and Claude Mythos 5“ - Anthropic

• “Claude Fable 5 – is this Mythos model worth the wait?“ [VIDEO] - How I AI

• “Nine Things About Claude Mythos 5 That Matter If You’re Not an Enterprise Customer“ - The Algorithmic Bridge

• “Vibe Check: Fable 5 Is the Best Coding Model in the World“ - Every

• “What it feels like to work with Mythos“ - One Useful Thing

🫵 Over to you…

Have you tried Fable 5 yet? If so, how much of a difference did you notice on your tasks and projects? Is my hot take way off? Is it way on? Am I rambling?

Leave a comment or drop me a line 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 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 my work and unlock cool perks, consider a paid subscription:

Subscribe now

Welcome back to the weekly AI news roundup.

See all 100+ paid subscriber bonuses

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

👩‍💻 AI releases

1. Google open-sourced Gemma 4 12B, a multimodal model that processes text, images, and audio natively in a single pass and can run on a home laptop.

2. H Company open-sourced Holo3.1, a family of AI agents that can click, type, and navigate apps on the web, desktop, and mobile while running on your own device.

3. Ideogram launched Ideogram 4.0, an image model with the most reliable text rendering of any open-weight release, free for non-commercial use.

4. JetBrains open-sourced Mellum2, a fast coding model that can act as a specialist in bigger multimodel pipelines to handle routing, summarization, and refactoring.

5. Krea released Krea 2 Turbo, a faster version of Krea 2 that can create high-quality images in about two seconds. (Try it for free.)

6. Microsoft rolled out its first in-house coding model, MAI-Code-1-Flash, directly inside GitHub Copilot, free and optimized for fast, lightweight coding tasks.

7. MiniMax open-sourced M3, a natively multimodal model with 1M context, frontier performance, and coding skills designed for long-running agent tasks.

8. Nous Research launched Hermes Desktop, a native app for installing, configuring, and chatting with Hermes Agent without using the terminal.

9. OpenAI news:

   1. ChatGPT can now send emails directly from a conversation, so you can draft, edit, and fire off a message without leaving the chat.

   2. Codex can now adapt to your specific job role thanks to plugin bundles for marketing, sales, support, and more with relevant apps and skills.

   3. Dreaming lets ChatGPT process your past chats in the background to build a persistent understanding of your preferences. (Paid US accounts only for now.)

10. Reve launched Reve 2.0, a 4K image generator that lets you control exactly where elements appear using visual layouts rather than just text prompts. (Try for free.)

11. Runway brought its Aleph 2.0 model to the API, letting developers make precise video edits while preserving the rest of the clip.

12. TwelveLabs launched Rodeo, an AI video copilot that lets you find and assemble footage using plain language without manual scrubbing.

13. xAI released Grok Imagine 1.5 Preview, which lets you turn still images into cinematic 720p videos with natural-language prompts.

🔬 AI research

1. GitHub announced a Copilot Desktop App in technical preview, letting developers manage coding agents, create secure sandboxes, and review changes.

2. Microsoft introduced Scout, an always-on Microsoft 365 agent that schedules meetings, flags stalled decisions, and learns your priorities over time.

3. Perplexity teased a hybrid inference system for Perplexity Computer that automatically splits AI tasks between your device and the cloud.

📖 AI resources

1. “EVA-Bench Data 2.0” [BENCHMARK]: a set of 121 tools and 213 real-world scenarios to test AI voice agents across different enterprise domains.

🔀 AI random

1. Anthropic confidentially filed a draft S-1 for a proposed IPO with the Securities and Exchange Commission. (But let’s hope it gets its house in order soon.)

🤦‍♂️ AI fail of the week

I don’t know what’s happening here, but we need this in the Olympics ASAP.

📹 Live & Learn #5: AI logo makers

I’m back with the Live & Learn concept, where I test and rate AI tools, like so:

• I pick 3-5 tools in the same category

• I test them live and rate them on 3-5 specific dimensions

• I end up with a leaderboard based on my obserations and ratings

This week, I looked at and ranked five AI logo generators:

In the next Live & Learn, I will be testing chatbots and standalone tools that let you make edits to images.

Join the session on Friday, June 12, at 2PM CET (8AM EST):

Live & Learn: AI Image Editing

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!

We’re just two-and-a-half days into June, with 8 “incidents” across multiple models.

As if that wasn’t enough, here’s some good ol’ Reddit bashing of the newest Opus 4.8:

Just a few months ago, Anthropic was everyone’s darling.

What happened?

January-February 2026: “QuitGPT”

At the start of the year, Anthropic’s Claude Code product was riding a wave of mainstream popularity.

Even reluctant laggards like yours truly finally jumped on the bandwagon.

Businesses started picking Anthropic over OpenAI for their first-time AI spend:

Opus 4.6 came out in early February and soon became the first model to ever be #1 across three separate Arena leaderboards (code, text, and search) at the same time:

By most measures, Anthropic was already doing pretty damn well.

And then came QuitGPT.

After Anthropic refused to budge in the standoff with the US Government over the use of its AI models for mass surveillance and fully autonomous weapons, the administration designated Anthropic a “supply chain risk” to discourage third parties from conducting any business with the company:

Shortly after, employees at Google and OpenAI signed a “We Will Not Be Divided” open letter in support of Anthropic’s stance (1,000+ signatures as of now).

Shortlier afterer, OpenAI’s CEO Sam Altman posted this:

All. Shit. Broke. Loose.

OpenAI instantly became the main villain in everyone’s “Dario vs. Government” “David vs. Goliath” story: Here was Anthropic standing by its moral principles only for OpenAI to swoop in like a blood-starved vulture and bite a big chunk out of that juicy government contract’s flesh.1

This fueled the already ongoing QuitGPT movement.

People started fleeing from ChatGPT to Claude in droves.

Anthropic dropped an “Import Memory” feature that made it super easy for users to switch to Claude from “other AI providers” (wink wink, nudge nudge).

By early March, Claude was the #1 app on the US Apple App Store.

And Anthropic lived happily ever after.

No, wait…

Subscribe now

March 2026: “Too big, too fast?”

The problem with having the most downloaded app on the App Store is that your app is downloaded by a whole lot of people.2

The problem with a whole lot of people trying to access your app at the same time is, well…

Yup.

On March 2, right after hitting that sweet #1 spot, Anthropic’s services couldn’t cope with the sudden influx of users and ran into compute constraints. (CEO Dario Amodei would later publicly acknowledge this.)

But hey, at least it was just one bad day, right?

Right?!

Oh, I see.

Not content with simply facing these ongoing bandwidth issues, Anthropic also quietly scored several own goals.

• March 4: Claude Code’s default reasoning effort was lowered from ”high“ to “medium,” leading to a performance drop that power users instantly noticed.

• March 26: A new bug made Claude "forget” prior turns in a conversation, which is less than ideal if you prefer working with chatbots that don’t have amnesia.

It certainly didn’t help that Anthropic wouldn’t acknowledge or address this until almost two months later in an April 23 postmortem.

Finally, Claude’s usage limits are way stricter than ChatGPT’s, especially on free plans. This was apparently not something the ex-ChatGPT crowd was prepared for:

It looked like the honeymoon was over before it truly began.

But not to worry: All Anthropic needed was a shiny new model, and all would be forgiven.

Cue dark foreshadowing music.

Subscribe now

April 2026: “Gaslightus-4.7”

Anthropic’s technical challenges continued well into April:

On forums like Reddit, the emerging mood could be loosely summarized as “Stop constantly shipping new goddamn features and just fix the goddamn basics, bro.”

Then, on April 16, Anthropic proudly released a brand new model: Opus 4.7.

Anthropic called it a “notable improvement on Opus 4.6 in advanced software engineering, with particular gains on the most difficult task.”

Users called it “a genuine regression,” “a serious regression,” “legendarily bad,” and—ominously—“Anthropic’s downfall.”

One of the posts referred to Opus 4.7 as “Gaslightus-4.7” because it tended to keep denying its own errors even when the user provided clear proof.

The name stuck because people felt it mirrored Anthropic’s own behavior. The company, faced with increasing reports of degraded performance, chose to play the “I know you are, but what am I?” card:

To add fuel to the fire, Anthropic shipped Opus 4.7 with a new tokenizer that effectively bumped usage costs by 20-30%.

It was around the same time that I started running into session limits more quickly and brought in Codex to work alongside Claude Code.

But Anthropic wasn’t quite done spraying bullets directly at its own feet.

On April 21, someone pointed out that Anthropic no longer listed “Claude Code” as a Pro plan feature on its pricing page. Seeing how Claude Code was a major reason so many people (including me) had a paid plan in the first place, this did not go over well at all.

Anthropic instantly backtracked, returned the old pricing page, and claimed it was just a small test on a subset of new users:

But all that did was feed the already growing “gaslighting” narrative.

After all the backlash grew into near-outrage, Anthropic reluctantly published the now-infamous post-mortem about all the performance issues, resulting in this delicious headline in The Register:

Anthropic should give masterclasses on rapidly squandering massive public goodwill.

But maybe not all was lost?

Subscribe now

May-June 2026: “More compute, less goodwill”

On May 6, Anthropic got some much-needed help from an unlikely source: Elon Musk’s SpaceX and its vast compute capacity. Thanks to this new deal, Anthropic could stabilize its uptime stats and raise rate limits for all paid users.

And on May 28, Anthropic released its latest model: Opus 4.8, specifically to address many of Opus 4.7’s blind spots.

But early reports from real-world users aren’t rosy. The short of it is that most are planning to stick to Opus 4.6 until Anthropic pries it from their cold, dead hands.

Visit any Anthropic-related subreddit, and you’ll find post after post complaining about general performance issues and Opus 4.8 in particular.

But that’s just anecdotal user reports, right?

Well, not exactly: While Anthropic reported better performance on handpicked benchmarks, several independent benchmarks hint at the real situation being a mixed bag at best.

LiveBench ranks Opus 4.6 (and even Sonnet 4.6) above both Opus 4.7 and 4.8 on agentic coding:

Arena has Opus 4.6 above Opus 4.7 on “Expert” text-to-text tasks (Opus 4.8 not ranked yet):

And SimpleBench puts Opus 4.6 in the lead out of all Opus models:

It almost feels like Anthropic is rushing model updates out the door to win back public sentiment, but it’s not quite getting the slam dunk it needs.

Now what?

On June 1, Anthropic submitted preliminary IPO paperwork to the SEC.

The company is currently valued at just under one trillion dollars.

On most metrics that matter to investors, Anthropic is doing just fine.

But if public “vibes” were a metric, they’d tell a different story.

The narrative playing out right now shows that an initial burst of goodwill isn’t enough to compensate for uptime issues, poor PR moves, and subpar product launches.

For Anthropic’s sake, we should hope it wins back the public’s favor and its status page starts logging fewer daily incidents than the screenshot in my intro.

Because you can’t IPO your way out of performance issues and eroding trust.

Thanks for reading!

If you enjoy my scribbles, 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

But for this Live & Learn, I wanted to test tools that either position themselves as dedicated AI logo makers or at least offer specialized templates/features for logos.

I narrowed the field to a mix of design tools with AI capabilities recently added, AI image tools with standalone logo templates, and an AI-first logo maker.

I gave all five the same prompt and scored them across the same five dimensions to keep the comparison as fair as possible.

Tools tested & scorecard

Read more

Welcome back to the weekly AI news roundup.

See all 100+ paid subscriber 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. Anthropic news:

   1. Claude Opus 4.8 brings incremental upgrades over 4.7 and several new features at the same cost. (But initial reactions are not positive.)

   2. Dynamic Workflows let Claude Code write orchestration scripts to run hundreds of parallel subagents to handle complex tasks.

2. ElevenLabs news:

   1. Dubbing v2 preserves emotional nuance when translating video content across languages by conditioning it on the original video. (Try for free.)

   2. Music v2 lets you create full songs that can switch genre mid-track with improved instrumentals, multilingual support, and richer vocals. (Try for free.)

3. Microsoft news:

   1. 365 Copilot is now more deeply embedded and loads faster inside Office apps like Excel and Word, so you can invoke it without breaking your flow.

   2. MAI-Image-2.5 is #3 on Arena’s text-to-image leaderboard and delivers sharp text and commercial-quality photography. (Try it free on Arena.)

4. OpenAI brought Codex computer use and remote control to Windows, so it can operate your desktop apps and continue coding sessions from the phone.

5. Perplexity made Computer available directly in Excel, Outlook, PowerPoint, and Word as native add-ins, so you can use it without leaving your Office apps.

6. Runway launched an MCP server that lets you create images and videos directly inside ChatGPT, Claude, Cursor, or Replit without leaving the app.

7. xAI launched Grok Build, its coding agent with automation tools and image/video generation, in early beta for SuperGrok and X Premium+ users.

🔬 AI research

1. Apple previewed a new Siri for iOS 27 powered by Google Gemini, with improved reasoning, persistent chat history, and a standalone app.

📖 AI resources

1. “DeepSWE” [BENCHMARK]: New “contamination-free” benchmark that measures coding agents on original, long-horizon software engineering tasks.

2. “How to evaluate AI agents (2026 edition)” [GUIDE]: practical guide to testing AI agents in production by Ben Hylak.

🔀 AI random

1. YouTube is making AI labels more prominent across the platform and adding automatic detection to help spot photorealistic AI content.

🤦‍♂️ AI fail of the week

“Hey, barkeep, got any of them self-filling bottles with self-making cocktails that spontaneously catches on fire?”

“Say less.”

📹 Live & Learn is back

During the next Live & Learn session, I’ll be testing and rating AI logo makers.

Join the session on Tuesday, June 2, at 2PM CET (8AM EST):

Live & Learn: AI Logo Makers

Let me know what would make the session extra useful by taking this quick survey:

Help shape "Live & Learn"

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

Thanks!

At the same time, I’ve consistently advocated for using AI actively as a sparring partner rather than passively as a substitute for your own thinking.

Much of that is just my nature: I think best when I get to bounce ideas off of others, and AI often provides a great sounding board.

But I also always felt that uncritically trusting AI without human oversight is a profoundly bad idea.

Well, it turns out that—plot twist—uncritically trusting AI without human oversight is a profoundly bad idea.

Who would’ve thought?

Today, I want to take a quick look at the concept of “cognitive surrender” and what you might be able to do about it.

What is “cognitive surrender”?

“Cognitive surrender” refers to uncritically accepting AI outputs in place of relying on your own squishy brain.

The term was coined earlier this year by Wharton’s Gideon Nave and Steven Shaw in a research paper titled “Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender.”

The study used a version of the cognitive reflection test and found that, even when AI output was wrong, almost 80% of participants still chose to go with AI answers without overriding them.

Man, 80% is, like…consults ChatGPT…almost half of all people!

So that’s, you know, not great.

But while the term “cognitive surrender” is relatively new, the risk of blind reliance on AI isn’t.

New name, old observations

Look, I don’t mean to toot my own horn, but…

…with that out of the way, I want to share a few posts where I warned against cognitive surrender well before that combination of words was used to describe it.

Back when ChatGPT first rolled out AI web search, I explicitly called out the risk of entrusting your research wholesale to a chatbot:

Here’s what I wrote:

And even if AI summaries are magically 100% accurate, what will be the long-term effect of us settling for condensed AI snippets instead of doing our own deep research?

My fear is that at least for a while, AI search will only amplify the Dunning-Kruger effect, giving us the illusion of knowledge by pacifying us with a seemingly authoritative filter between ourselves and the source.

But beyond just being a wise guy who points out the “wrong” way to do stuff, I also shared my ideas on what the “right” way might look like:

• In “The Skeptical Writer’s Guide to AI,” I showed how writers could use AI while staying in the driver’s seat and not giving up what matters.

• In “How to Level Up Your Job Hunt With AI,” my main message was to use AI to make yourself a better candidate instead of automating the application process.

• In “How to Master New Topics with AI,” I covered a few approaches to turning AI into a tool that helps you learn instead of doing the thinking for you.

Hopefully, reading these posts will help you embrace AI while keeping your critical thinking intact.

Also, if you’re specifically trying to learn a new subject or understand a concept, remember to switch on the dedicated “study” mode in your chatbot of choice:

Are there other things you can do to avoid surrendering to AI?

Well…

Subscribe now

5 habits that help prevent cognitive surrender

Cognitive surrender isn’t always a single clear instance. It may well creep up over time as you gradually trust AI more and more at the cost of exercising your own judgment.

One way to fix this is to read every Why Try AI post ever written.

If, for some odd reason, that’s not on your bucket list, here are a few other ways you can train yourself to stop deferring to AI.

Habit #1: Form your own opinion first

Before you ever ask AI for input, formulate your own take.

It’s okay if it’s just a few scribbled notes or a stream-of-consciousness rant you dictate into a microphone. The important part is doing your own legwork instead of letting AI do it for you.

If you know where you stand, you won’t accept AI output without pushback.

Habit #2: Ask for a challenge, not validation

Instead of turning to a chatbot with an open-ended “What do you think?”, try to intentionally invite friction by asking AI to poke holes in your assumptions, point out what you might have missed, find counterarguments, and so on.

AI chatbots are notoriously sycophantic, but asking one to stress test your ideas nudges it out of this default behavior. I shared seven anti-sycophancy tips here:

Habit #3: Always check the source

I mean, this has been a good rule of thumb even in the pre-AI era.

If a chatbot cites a verifiable fact or figure, go to the source. Does the source actually say what AI claimed, or was it just a confident hallucination? What’s the context around the claim, and does it change the picture AI presented?

Understanding where a piece of information came from and how it was obtained is often as important as the information itself.

Habit #4: Describe things in your own words

After you get a finished product or output from AI, pretend you’re giving your imaginary friend1 an executive summary.

If you can easily find the right words, your brain has been engaged in the process.

But if you struggle, that’s a good sign you might be deferring to AI a tad too much. Revisit the task and see if you can internalize what AI’s been doing.

Habit #5: Don’t delegate what truly matters

Here’s how I try to approach AI: If a task is my primary deliverable, my core area of expertise, or simply something I’m passionate about, I never let AI do it for me. Sure, I might ask AI for a second opinion, but the bulk of the effort and output will come from me.

For instance, no matter how good AI gets at writing, I’ll still be stringing my own words together. That’s because I like writing. I get a kick out of sprinkling in silly asides and jokes and knowing nods to my audience. (You know who you are, Fred, wink wink.)

But I’ll happily turn to AI for filler videos or images or other visuals to supplement my writing, because a) I suck at most visual arts and b) it’s not a skillset I want to develop or consider critical to my identity.

You might be the exact opposite of me: a passionate cartoonist who won’t be caught dead using AI to create images, but who’ll happily hand tedious writing tasks to the first chatbot you come across. And your name is Leinad.

The point is that you should protect the work that makes you who you are while letting AI handle the busywork, however you define that in your role.

That way, you won’t have to watch your creativity and thinking muscles atrophy.

Here’s a quick reference sheet:

For an additional perspective, check out this wonderful TED Talk by Advait Sarkar:

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 AI news roundup.

See all 100+ paid subscriber 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 released Qwen3.7-Max, its new flagship agent model built for coding, workflow automation, and long-horizon tasks. (Try for free at Qwen Chat.)

2. Amazon launched Alexa Podcasts, so you can request a custom audio episode on any topic that Alexa researches, writes, and narrates. (Alexa+ users only.)

3. Cohere open-sourced Command A+, a mixture-of-experts agentic reasoning model that can run on just two H100 GPUs. (Try for free on Hugging Face.)

4. Cursor launched Composer 2.5, a coding agent that matches frontier model performance at a fraction of the cost.

5. Google news (mainly from I/O 2026):

   1. AI Mode was upgraded with Gemini 3.5 Flash and now includes search agents and can build mini-apps to handle tougher questions and tasks.

   2. Antigravity 2.0 is a new standalone desktop app that can orchestrate many AI coding agents at the same time and schedule background tasks.

   3. Gemini 3.5 Flash is the company’s best-in-class model for coding and long-horizon agent tasks that’s also four times faster than frontier alternatives.

   4. Gemini Omni Flash can create and edit videos from any combination of audio, images, text, and video inputs.

   5. Gemini Spark is a 24/7 personal AI agent that handles multi-step tasks across your Calendar, Docs, Drive, Gmail, Sheets, and more.

   6. Pomelli gained agentic features that can help businesses create brand identities, brand books, and full websites in just a few clicks.

   7. Project Genie can now use Google Street View to let you explore an AI-generated version of any real-world location. (Gemini Ultra users for now.)

   8. Stitch Agent lets you build, stream, and steer app designs in real time from text, voice, code, or existing design files. (Try for free.)

   9. SynthID expanded its AI watermarking across Chrome, Gemini, Google Search, and Pixel to help people spot when content was made with AI.

6. Granola launched Briefs that generate a pre-meeting summary of what was discussed last time and what to focus on just before you join.

7. Ideogram launched an MCP server to let AI chatbots and agents generate and edit images directly in the chat.

8. Krea released LoRA fine-tuning for Krea 2 in beta, so you can train the model on your own images to lock in a specific style, character, or object.

9. Leonardo AI launched 3D Generation, which can turn a prompt or image into an exportable .glb model without leaving the platform.

10. Microsoft launched an improved Bing Image Search that organizes image results into AI-labeled groups that help you find what you’re after faster.

11. OpenAI news:

    1. Appshots lets Mac users send any open app window straight into Codex without copying anything manually.

    2. ChatGPT for PowerPoint lets you generate and edit Microsoft PowerPoint slides using natural language. (Try for free during global beta)

    3. Verify is a new tool that lets you check whether an image was made or edited by OpenAI’s models. (Try it for free.)

12. Stability AI open-sourced the Stable Audio 3.0 family of music models that can generate 6-minute tracks and were trained on fully licensed data. (Try it for free.)

🔬 AI research

1. Apple previewed AI-powered accessibility updates like Vision Pro eye-tracking wheelchair controls, natural language Voice Control, and more.

2. Figma announced its Design Agent that works inside your file, follows your design system, and makes edits directly on the canvas. (Request early access.)

3. Google news:

   1. AI Studio is adding a mobile app so you can start AI projects on your phone and continue on your desktop. (Pre-register for the app now.)

   2. Universal Cart will let you add items via Gmail, Gemini, Google Search, and YouTube into an AI cart with automatic price alerts and stock notifications.

4. Microsoft previewed MagenticLite, a lightweight multi-agent framework optimized to run browser and file tasks with small models.

5. Spotify news:

   • Fan-made covers will let Premium subscribers create AI-generated covers and remixes of songs from participating Universal Music artists.

   • Personal Podcasts can create AI audio episodes on any topic from a prompt, PDF, or link on a recurring schedule. (Premium subscribers only.)

📖 AI resources

1. “Two Rival Bets on AGI: Google I/O Highlights” [VIDEO]: fascinating look at the different pathways to AGI pursued by different labs, from AI Explained.

🔀 AI random

1. Anthropic hired Andrej Karpathy (ex-Tesla and OpenAI co-founder) to work on Claude pre-training and speeding up frontier model research.

2. OpenAI had its reasoning model autonomously disprove an 80-year-old “unit distance” geometry problem, with results verified by Fields medalist Tim Gower.

🤦‍♂️ AI fail of the week

I wanted distorted, surrealist version of a person reflected in the mirror, following their motions. This…is not that.

📹 Live & Learn: What shall I tackle next?

After a brief break, I’m planning to return to my weekly Live & Learn sessions again.

Let me know what types of tools you’d like me to test:

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

I argued that storing my knowledge in an Obsidian vault would make it easier for future agents to surface background details and pick up ongoing projects. And that’s exactly what happened when I made Claude Code and Codex work together.

But my Obsidian vault has evolved into more than just a reference library.

It now functions as a task inbox and project manager, too.

So today, I’d like to walk you through my Claude Code setup and show you where Obsidian fits into the picture.

(I also have a bonus goodie that can help you build out a similar routine.)

Let’s roll!

A typical day in my life with Claude Code might involve four different activities, all tied to the context stored in my Obsidian vault.

Let’s look at them in vaguely chronological order.

✍️ 1. Capture: Brain Dump.md + my phone

Inside my Obsidian vault, there’s an Inbox folder with a Markdown file called “Brain Dump” where I can log just about anything I want to tackle with Claude.

In addition to my laptop, I also have the Obsidian Android app installed on my phone.

As I go through the day and batshit crazy genius ideas pop into my head, I can type (or dictate) them into the Brain Dump, which has two distinct sections:

• Brain Dump: This is the classic note-capture for any ideas and plans.

• Learning Queue: This one’s for articles, YouTube videos, and other sources related to how Claude Code works (tutorials, best practices, tips, etc). I simply copy-paste the link to the article or video in question, one per line.

At the end of the day, the file might look something like this:

Because I use Syncthing to sync the Obsidian vault between my devices, the local Brain Dump file is automatically updated when I turn on my laptop.

✅ 2. Triage: New Claude Code session

Whenever I start a fresh Claude Code chat, it triggers a special SessionStart hook.

This asks Claude to do two things:

1. Check the Brain Dump file for any new notes or learning URLs I might’ve added.

2. Pick a high-priority project from our work-in-progress list to tackle (see #3).

This happens automatically, without me having to request it, like so:

Claude surfaces all my notes, offers its initial reaction, and asks me what we want to tackle. From here, two things typically happen.

First, Claude and I go through the Brain Dump items together.

Second, Claude processes the learning queue and extracts insights we can apply to our setup.

Let’s look at each of those a bit more closely.

Work through the Brain Dump (together)

I discuss each note with Claude and provide additional context.

We go back and forth to decide what should happen with that particular item. The most common outcomes are:

• We resolve the note together in the same session (for quick tasks/ideas).

• We turn it into a bigger project and log it in our Projects.md file (see #3).

• We decide it’s not worth focusing on and simply clear that item from the list.

Here’s how that might look:

After I reluctantly abandon my dreams of Beanie Baby riches, Claude can move on to the Learning Queue.

Process the Learning Queue (Claude)

Every URL in the queue triggers my learning-extractor skill that does the following:

• Fetches article content (or the transcript in the case of YouTube videos)

• Analyzes the article or video and extracts key insights

• Checks these against our current Claude Code workspace

• Splits the findings into three buckets:

  • Already doing: Things already implemented in our setup.

  • Could adopt: Things we could implement right away if we wanted to.

  • Watch list: Interesting findings we can’t act on yet because of missing prerequisites, plan limitations, or priorities.

Here’s an example for the claudefa.st/blog/guide/development/routines-guide article:

As you can see, most Brain Dump ideas and Learning Queue items will also end up touching at least one of two files in the Obsidian vault:

• Insights.md: Stores context about the way I work, relevant tools and procedures we came across, and anything else relevant to my setup. Claude Code can reference this file on the fly whenever we work on a related project or task.

• Projects.md: This is our Obsidian version of a Trello board where Claude Code and I log projects, prioritize them, and track their progress. I don’t really touch this file directly. Instead, Claude brings up relevant items in any chat and updates projects accordingly.

And hey, what a suspiciously convenient segue to my very next point…

Subscribe now

🛠️ 3. Build: Grab an existing project to work on

If I still want to keep going after clearing the Brain Dump, Claude will surface a project on our list based on its priority and status:

As you can see, Claude and I decided to work with the concept of “bites”: small sub-tasks in a project that can be realistically knocked off in a single session.

That’s not an accident.

Quite simply, I ended up with a long list of big projects that never moved anywhere, because they felt too overwhelming to tackle in the middle of an ongoing session. So I asked Claude to split each project up into self-contained tasks.

This way, I can complete just one minor task and still feel like the project is moving forward. Once the task is finished, Claude updates the project status in Obsidian, so it’s ready to be picked up in a future session.

🧠 4. Enrich: Save and organize new insights

Each of the first three steps will usually generate new knowledge along the way.

Brain Dump notes get turned into detailed projects, learning queue surfaces new approaches to try, my in-chat input gives Claude additional context to log, and so on.

Whenever something like that happens, Claude automatically runs a quick “fanout” process: identifying where a piece of knowledge belongs and filing it accordingly.

These learnings can end up in several places:

• New notes: Whenever a concept (person, company, tool, etc.) shows up in at least three other spots in the vault, it becomes a candidate for a standalone note. Claude “upgrades” the concept and moves all existing details about it into a new, dedicated note.

• Existing notes: If there’s already a note for the concept, additional info about it gathered from the session gets filed there.

• Related notes: If Claude spots a relationship between separate notes—like if it learns that Andrej Karpathy now works for Anthropic—it’ll update the separate “Karpathy” and “Anthropic” notes and create a wiki connection between them.

• Workspace files: If a detail is related to my work with Claude Code, it’ll be added to the relevant section of something like Projects.md or Insights.md (or both).

Here’s how that might look after we extract learnings from that “Claude Routines” guide:

Each edit might be minor on the surface (e.g. a single line added to an existing note), but multiply that across dozens of sessions and fanout runs, and you end up with a constantly growing, interconnected knowledge web.

🔁 The self-updating Obsidian loop

Let’s recap what just happened.

I saved a few thoughts and links on my phone. Claude triaged them at session start and pulled relevant context from existing notes and projects into the chat.

When we were done working, Claude collected and saved new insights, enriching the knowledge base in the process.

It’s a closed loop: Captured ideas trigger the triage, which relies on past notes for context, which feeds into the session’s work, which generates new insights, which get saved as projects and/or new notes, which serve as better context for future sessions.

Rinse. Repeat.

At the core of this lies the Obsidian vault that serves as the primary anchor for idea capture, projects, insights, notes, and everything else.

The cool part is that I don’t have to actively maintain this system. Claude does this on its own as we work. Every session leaves the vault a little more connected, and every future session kicks off from a stronger foundation as a result.

Pretty damn cool, eh?

Subscribe now

Want to build your own agent control center?

What I showed you is my setup for Obsidian and Claude Code, but the premise works with other ways of capturing notes and any AI agents that can modify them.

To build your own version, you’ll need four components:

1. Capture point: A file where you save your ideas, thoughts, and research links.

2. Startup check: Instructions that tell the agent what to scan at session start.

3. Project tracker: Dashboard, sheet, .md file, etc. that keeps track of your projects.

4. “Fanout” routine: Instructions or skills that update notes with new info.

Everything you saw above runs on my custom variations of these four elements.

You now have all the ingredients to build out a control center of your own.

If you want to skip the manual setup, I put together a starter kit that does the heavy lifting for you. You paste one prompt, it scans your workspace, asks you questions, and creates a version tailored to the way you work:

If that’s the kind of thing that tickles your fancy, find it on the other side of the paywall.

Read more

Welcome back to the weekly AI news roundup.

Check out the "Sunday Bonus" directory

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. Amazon launched Alexa for Shopping that can research products, compare prices, track deals, and automate routine purchases.

2. Anthropic news:

   1. Agent View lets you manage all your Claude Code sessions from a single dashboard to see which ones are running, need input, or are completed.

   2. Claude for Legal comes with 20+ MCP connectors and 12 plugins tailored to specific legal work and practice areas.

   3. Claude for Small Business includes 15 ready-to-run workflows for finance, HR, operations, and sales with connections to Canva, PayPal, and more.

3. Krea launched Krea 2 image model with style transfer that lets you match any aesthetic from reference images and moodboards.

4. Meta news:

   1. Incognito Chat lets you have fully private AI conversations on WhatsApp with messages that don’t get saved or seen by anyone, including Meta.

   2. Muse Spark went from research to general availability and is now accessible across all of Meta’s apps (Facebook, Instagram, Messenger, and WhatsApp).

5. Microsoft added new Copilot features to Edge on desktop and mobile, including long-term memory, multi-tab reasoning, and study tools like flashcards.

6. OpenAI news:

   1. Codex is rolling out to the ChatGPT app on iOS and Android, letting you kick off coding tasks and other long-running projects from your phone.

   2. Personal Finance connects ChatGPT to 12,000+ financial institutions, so you can ask questions about your spending and get live dashboards in the app.

🔬 AI research

1. Google news:

   1. Gemini Intelligence is coming to Android this summer with Gemini built into Chrome, multi-step task automation, and intelligent speech-to-text.

   2. Gemini Spark is an always-on AI agent that runs in the background to handle your daily tasks (expected to be announced at Google I/O).

   3. Magic Pointer will turn your mouse cursor into a context-aware mini AI agent that intuitively understands your requests based on where you’re pointing.

2. Thinking Machines Lab previewed interaction models that can respond, interrupt, and translate in real time across audio, video, and text.

3. xAI announced Grok Build, a coding agent that can write files, run commands, and build apps from natural language. (Early beta for SuperGrok Heavy users.)

📖 AI resources

1. “AI IQ” [BENCHMARK]: a new platform that scores frontier AI models on an IQ curve, with cost comparisons, intelligence trends, and live rankings.

🤦‍♂️ AI fail of the week

This clip of me playing chess against a robot is practically flawless, but an eagle-eyed Grandmaster might be able to spot a few subtle inconsistencies.

📹 My Cozora image-to-video workshop

On May 7, I had the pleasure of joining of Cozora for a one-and-a-half-hour practical session on image-to-video models.

I shared hands-on examples of using images with AI video, showcased different techniques, and discussed their practical applications.

It was a lot of fun and jam-packed with actionable tips. I even threw in a take-home reference sheet with key takeaways.

Paid subscribers can watch the full session here:

Watch the Workshop

If you’re not already a member, note that all of my subscribers are eligible for discounts on Cozora memberships:

• Free subscribers get 10% off

• Paid subscribers get 50% off

In related news: I expect to get back to my Live & Learn sessions the week after next, so if you have input on future topics, format, and so on, please share it here:

Help shape "Live & Learn"

Image-to-Video AI: Hands-On Masterclass (Daniel Nest)

Most people have seen AI-generated videos online, and most of what they’ve seen is slop. That reputation is both earned and misleading. It’s earned because the easiest thing to do with these tools is to produce something low-effort and forgettable. It’s misleading because the tools themselves are far more capable than their worst output suggests.

This session was a deep dive into the practical, purposeful side of image-to-video AI —the side that doesn’t get much attention. My guest, Daniel Nest, is the creator of Why Try AI (16,000+ subscribers), a passionate explorer who has been hands-on with these tools since the early image model days of 2022. He came with demos, frameworks, and a genuine nerd’s enthusiasm for showing what’s actually possible when you approach these tools thoughtfully.

Whether you’re a content creator, marketer, writer, or filmmaker, the techniques Daniel shared have real-world applications. And the core insight running through all of it is deceptively simple: the smarter you are about what you give the model, the less work you have to do — and the better the results.

If you weren’t there, the replay is worth your time.

During the class, we:

• Toured the image-to-video tool landscape, from beginner to advanced

• Learned why Google Flow is the best starting point for newcomers

• Explored the 5S Framework for building effective video prompts

• Saw how image references can replace complex text descriptions

• Demonstrated the “starting frame” technique for precise scene control

• Used first and last frames to let the model fill the narrative gap

• Discovered how annotations and scribbles direct complex on-screen action

• Explored character and object reference for consistent brand storytelling

• Learned multi-panel storyboarding to generate narrative-driven clips

• Discussed where image-to-video AI is heading — and its real ceiling

Bonus: Image-to-Video Reference Sheet

Daniel has packed all of the insights and tips into this reference guide that includes:

• The 5S Framework for prompting image models

• The 6 image-to-video techniques and use cases

• List of today’s best AI image models & multi-model platforms

• Image model benchmarks & live leaderboards

• Different prompting approaches & when to use them

• Camera movement vocabulary

• And useful references for further reading

Read more

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

I may have written about Claude Code once or twice before.

Perhaps even thrice.

Okay, fine, five times, but who’s counting? Oh, am I counting? Get out of here!

The point is: Ever since I first tried Claude Code back in January, it’s been my daily go-to driver for most tasks.

My entire workspace was built around Claude Code. I had multiple Claude skills running, used session-start hooks to make Claude auto-check for new tasks, and hooked Claude Code up to Obsidian as its primary knowledge base.

In short, I was all-in on Claude Code.

But two recent developments made me re-evaluate things:

1. Anthropic had a patchy stretch with a lot of downtime, launched Opus 4.7 to overwhelmingly negative reactions, and introduced new tokenization magic that meant I started hitting usage limits way faster than before.

2. OpenAI came out swinging with the launch of GPT-5.5, a streamlined Codex app for Windows, and a best-in-class GPT Image 2 model.

Suddenly, Codex was once again an attractive local agent option.

I knew I wanted to give Codex a shot, but I also didn’t want to abandon my now-familiar Claude Code routine.

So I wondered: “Can I make Claude Code and Codex work on the same workspace with little friction and clean task handovers?”

It turns out, I could.

Why use multiple AI agents?

Remember how I once said I wanted to create context to use with any future agent?

“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!’”

Well, for some odd reason, I always pictured this as an all-or-nothing decision.

In my mind, the day I decided to use a new agent, I’d just switch over completely.

But….like….why?!

Why not simply have two (or more) agents working in parallel?

Why not indeed.

After running my two-agent setup for a while, I’m happy to report that it comes with a bunch of real benefits:

1. Higher usage limits: My $20/month Claude Pro plan was no longer enough. If I were to stick to pure Claude, I’d have to either upgrade to the $100 Max plan or find a hacky and friction-y way to use two separate Claude Pro accounts. By plugging in a $20 ChatGPT Plus plan, I now spend $40 on two agents and get plenty of extra usage. If Claude hits a limit, I switch to Codex and keep going.

2. Three heads are better than two: I like having two sparring partners. Between my own judgment and separate agents with different training data, reasoning patterns, and blind spots, every decision gets more eyes on it. Each model usually notices something the other one had missed, and the final result is better for it.

3. Complementary skillsets: As I’ll explain in a bit, there are certain areas where Claude shines and others where Codex is the better choice. By leaning on their individual capabilities and strengths, I get a more well-rounded setup.

4. Future-proof workspace: The beauty of my current setup is that it’s not exclusive to Claude and Codex. I have a truly agent-agnostic system that lets me use any number of them. Also, there’s no lock-in: If a new agentic model comes out, I can plug it right in and pick up where I left off.

How I use Claude Code and Codex in tandem

While the setup is designed to work with any agent(s), my experience comes from using specifically Claude Code and Codex.

Here are my observations about the two:

Generally speaking, I still turn to Claude Code when I need to kick off a bigger project. It just feels better for high-level conceptual discussions and planning. I also find that Claude Code designs better HTML layouts and pages out of the box.

Codex has more of a “workhorse” vibe: reliable and fast. It also has native image-generation capabilities thanks to GPT Image 2. But I noticed that Codex is sometimes too eager to just run with the task instead of pausing to ask me questions or digging a bit deeper.

Having said that, the line isn’t quite as black-and-white as I make it sound.

In many cases, I’ll ask Codex for a second opinion and go back and forth between the two agents several times to flesh out a concept or plan.

Subscribe now

Two-agent workflow: Hands-on example

Let’s look at a not-so-hypothetical scenario of me redesigning the current Sunday Bonus Directory to incorporate all of my paid subscriber perks. (Paid subscribers: Stay tuned for the outcome of this process soon.)

I could boot up Claude Code and kick the project off:

Note how little in-chat background Claude needs because it already has access to my entire workspace, the Sunday Bonus Directory itself, and additional context files.

I then answer the six preliminary questions, and Claude Code comes up with a plan.

But I want Codex to weigh in, so I ask Claude for a handoff:

Yup, just one word.

Then I open Codex and ask for its thoughts:

Again, Codex doesn’t ask me for extra context because Claude’s handoff document and my workspace give it everything it needs.

I then request Codex to provide its own plan:

If I want to keep going back and forth, I can ask Codex to hand off…

…and then get Claude to review in my ongoing chat:

Normally, I’d be way more involved and opinionated throughout this process, and I wouldn’t go back and forth so much at this “early discussion” phase. I’d likely first switch to Codex when Claude and I had at least designed the initial architecture.

But I wanted to demonstrate just how frictionless things can be once the multi-agent system is in place.

Now let me show you how you can set up your own version of this.

How the setup works

Read more

Welcome back to the weekly AI news roundup.

Check out the "Sunday Bonus" directory

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. Adobe launched a Productivity Agent in Acrobat Express that turns PDFs into shareable AI workspaces you can interact with.

2. Anthropic news:

   1. Agent Templates for financial services give teams 10 ready-to-run Claude workflows to help with pitch prep, KYC screening, and more.

   2. Claude Code rate limits have doubled, and there are no longer peak-hour limitations for Pro and Max accounts.

   3. Claude for Microsoft 365 is out for all paid plans, bringing Claude into Excel, PowerPoint, and Word (as well as Outlook in public beta).

   4. Managed Agents now have Dreaming (automatic memory review between sessions), Multiagent Orchestration, and Outcomes (working against a rubric).

3. Cursor introduced /orchestrate, a new skill that spawns specialist subagents to handle complex tasks in parallel.

4. ElevenLabs added a Studio Agent to ElevenCreative: an AI collaborator embedded in the audio/video timeline that lets you create without switching tools.

5. Google news:

   1. AI search results will now quote first-hand insights from Reddit, forums, and social media, featuring creator handles and clickable source links.

   2. Gemini 3.1 Flash-Lite is generally available as the fastest, most cost-efficient model in the Gemini 3 family, designed for ultra-low latency tasks.

6. LTX Studio launched Flows, a node-based canvas that lets you design and run visual generation workflows in batches.

7. OpenAI news:

   1. GPT-5.5 Instant in ChatGPT comes with fewer hallucinations, better image analysis, and clear explanations of what memories shaped its response.

   2. Codex can now work directly in Chrome on macOS and Windows, letting you run parallel coding agents across multiple browser tabs.

8. Perplexity news:

   1. Computer for Professional Finance connects to licensed data providers like Morningstar and can run pre-built analyst workflows in Microsoft Teams.

   2. Personal Computer is now out as a native macOS app that can run agentic tasks across local files, Mac applications, the web, and 400+ connectors.

   3. Premium Health Sources will pull peer-reviewed clinical journals into Perplexity answers.

9. Spotify launched a beta AI personal podcast feature that creates and saves daily briefings and learning content directly to your library.

10. Unity launched Unity AI in open beta, letting developers generate scenes from images, write scripts, and create placeholder assets without leaving the editor.

11. xAI launched Grok Voice Mode for Apple CarPlay, letting you talk to Grok hands-free in your car.

🔬 AI research

1. Subquadratic introduced SubQ, an AI model that can read long documents in one go without having to split them into smaller chunks. (Request early access.)

📖 AI resources

1. "Turning Claude’s Thoughts Into Text" [RESEARCH]: Anthropic's new study on helping to decode Claude’s internal “activations” into plain text.

🤦‍♂️ AI fail of the week

My prompt was: “Two men tumble out of a dingy kitchen’s rear exit, arguing and fighting.” At no point did I request a door-glitching giant in a dissolving tank top.

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

A version of this article first appeared as a guest post for AI Supremacy. Since the article was first written, Genspark has released Workspace 3.0 and Workspace 4.0. But the interview and insights about the company’s long-term direction still stand.

Mere months after its Series B funding and “unicorn” status, Genspark launched the latest iteration of its agentic concept: AI Workspace 2.0.

AI Workspace 2.0 brings improvements to several existing agents and as well as new features:

New launches:

• Speakly: A new dictation assistant (think Wispr Flow and Superwhisper) that you can download for macOS or Windows. It lets you interact with all of Genspark’s agents and tools via voice instead of typing.

• AI Music Agent that can create custom music (think Suno).

• AI Audio Agent that does voiceovers and narration (think ElevenLabs).

Upgrades:

• AI Inbox now supports automated workflows that perform specific actions like creating daily inbox digests, interacting with external messaging platforms like Slack, analyzing social media performance, etc

• AI Creative Slides, AI Image Agent, and AI Video Agent are all more capable and incorporate the improved powers of newer, better underlying models.

While the under-the-hood upgrades and new music and audio agents are neat, what Genspark is leaning heavily into is Speakly and the promise of hands-free agentic work. Simply say what you need done, and Genspark’s Super Agent does it.

At least, that’s the idea.

To better understand Genspark’s vision, I tested the three new features and also conducted a written interview with the company’s COO, Wen Sang.

Read on to find out if Genspark can be more than the sum of its parts.

Testing the new tools

Let’s start with my hands-on tests and demos of the three new features.

1. Speakly: The voice agent

On the surface, Speakly is yet another voice dictation app. You speak into the microphone, and text comes out on the other end.

But what makes Speakly a great fit for Genspark’s infrastructure is the “agent” shortcut that sends your spoken request directly to Genspark’s Super Agent and the baked-in, intelligent processing of spoken input.

I demonstrate Speakly’s main features in this 7–minute hands-on video:

Watch it to learn:

• How regular voice dictation works

• How Speakly can auto-correct filler words and backtracking

• Agent mode: Send a request or task to Genspark Super Agent from any screen

• Translation mode: Speak in any language (or combination of languages), get English text.

• Custom modes: From “Buzzwords” mode to “Twitter” mode, Genspark can process and rework your speech into any style or format. You can customize this to your needs.

If you want to try Speakly for yourself, grab it here:

Download Speakly for macOS or Windows

2. AI Music: From requests to music tracks

If you have toyed with AI music sites like Suno or Udio, you already have a good idea about how this works. The agent can use different third-party AI music models to execute your task:

But because Genspark agents can coordinate their work with other agents, you can have more elaborate requests that require additional pre-processing. So I tried this:

Prompt: Watch this video about the launch of Genspark AI Workspace 2.0 and create a grand soundtrack worthy of this announcement: youtube.com/watch?v=KtOZNvFhV8c

The AI Music agent understood that it first needed to analyze the YouTube launch video before coming up with an appropriate soundtrack, so it invoked several additional tools:

After analyzing the nature of the announcement, it came up with a detailed second-by-second soundtrack structure:

All of this happened before any music was ever generated. Then the agent produced the soundtrack, and here’s the result:

Listen here

Now, the music might not exactly be your cup of tea.

But I find the multi-agent workflow that gets involved in the creation process to be a good reflection of Genspark’s grand “orchestration” vision that we’ll see below.

3. AI Audio: Create voice outputs like narration or podcasts

This one is similar to the AI Music agent, but for voice-based outputs using models by Google, ElevenLabs, MiniMax, and so on:

Just as before, I can submit requests that require additional analysis, without providing a script or dialogue directly:

Prompt: Watch this video about the launch of Genspark AI Workspace 2.0 and create a short podcast with two hosts: A skeptical woman who doesn’t buy the hype and an enthusiastic man who is excited about it: youtube.com/watch?v=KtOZNvFhV8c

Once again, the agent requests additional research from fellow agents and puts together a debate-style podcast:

Here’s the podcast itself:

Listen here

I think Genspark handled this one very nicely. We have two distinct voices and personalities discussing relevant takes from two sides of the same launch.

Q&A with Genspark COO Wen Sang

While I could certainly appreciate the sheer amount of new features Genspark has shipped in a single year, I wanted a better picture of what their grand vision was beyond “a convenient collection of third-party AI models and tools.”

Fortunately, Genspark gave me the opportunity to submit interview questions to its COO, Wen Sang. Below are my eight questions and Wen’s responses in their entirety.

Question 1:

Wen’s answer:

Most voice tools end at transcription. Speakly is designed for execution. In internal usage, voice-driven workflows are roughly 3-4x faster than typing because users move directly from intent to action instead of stopping at text. Speakly does not just format what you say. It can trigger agents, initiate workflows, and hand off work across documents, inbox, and media tools. That makes it a control layer for work, not a convenience feature.

Question 2:

Wen’s answer:

Email on autopilot is about reducing manual effort. With AI Inbox, users automate repetitive tasks like categorization, bulk cleanup, unsubscribing, and templated replies using custom agents. Early users are seeing a meaningful reduction in routine inbox work, often cutting manual email handling by around 30 to 50%. The system focuses on execution in the background so users can stay focused on decisions that actually matter.

Question 3:

Wen’s answer:

That is a fair description. The Super Agent acts as the orchestration layer. It interprets intent, plans the work, selects the appropriate agents, and coordinates execution across tools like slides, inbox, media, and voice. A single request often results in multiple agent calls running in parallel. The key point is that users do not manage those steps. They describe the outcome and review the result.

Question 4:

Wen’s answer:

The focus is convergence. Shared memory, shared assets, and shared intent across the system. Workspace 2.0 reduces what we call in-between work, the manual steps that sit between an idea and a deliverable that is ready to ship. Users are completing end-to-end workflows in significantly fewer steps, often replacing several disconnected tools with a single continuous flow.

Question 5:

If we look at the individual components in Genspark’s workspace—AI Slides, AI Sheets, image and video generation, and so on—all of them have standalone competitors. What’s Genspark’s “secret sauce” or integration layer that makes this collection more than the sum of its parts?

Wen’s answer:

The secret sauce is multi-agent orchestration. The integration layer is not just shared UI or shared data. It is a system where multiple specialized agents plan, coordinate, and execute work together across slides, sheets, inbox, media, and voice.

All of these tools operate on shared context and persistent memory, but the real value comes from how agents hand work off to each other automatically. Outputs are not endpoints. A presentation, an email, or a media asset immediately becomes input for the next agent and the next step. That is what turns a set of individual AI tools into a single system that can reliably execute end-to-end work.

Question 6:

Your current pricing plans closely mirror similar offerings from OpenAI, Google, Anthropic, and so on: The free plan, the mid-tier $20 plan, and the $200 plan. In your view, what makes Genspark a better competitive offering?

Wen’s answer:

At the same price points, Genspark gives users access to more than 70 models and tools across providers, including OpenAI, Anthropic, Google, and others. Instead of paying separately for each ecosystem, customers get a single workspace that selects the right model for the task automatically. The value is not just access, it is orchestration. Users do not need to decide which model to use or how to combine them. The system handles that complexity and routes work to the most appropriate model based on the task, context, and desired output. As teams move from experimenting with AI to relying on it daily, simplicity and flexibility become a clear advantage.

Question 7:

Genspark has a model-agnostic approach: You quickly incorporate fresh third-party tools into your ecosystem, which works for customers looking for a one-stop shop. What is your true ‘moat’ when it comes to these commoditized models? Is it the Super Agent orchestration layer, the workflow integrations, or something else?

Wen’s answer:

The moat is orchestration at scale. Models are becoming interchangeable. What is difficult to replicate is a system that can reliably plan, execute, and verify work across multiple agents and tools in production. That orchestration layer, combined with workflow depth and shared context, is where our defensibility lives.

Question 8:

Wen’s answer:

Surpassing $155 million in ARR in such a short time confirms that teams are ready to move beyond assistance into autonomy, but we see this as just the beginning. Our goal is to reach $1 billion in ARR in 2026, and we don’t want Genspark to be known simply as a tool that finishes work. We want it to be known as the platform that unlocks human potential by moving AI from reactive responses to proactive execution.

At a macro level, the future we’re building is about amplifying human capability, not replacing it. When knowledge workers can offload repetitive tasks and focus on judgment, creativity, and strategy, their impact goes up. We see a world where describing an outcome reliably results in production-ready deliverables, where AI agencies extend what individuals and teams can accomplish, and where organizations redefine leadership around human-AI collaboration, not human-only effort.

If we execute as planned, Genspark should be known as the operating system of intent-driven work, the place where AI does not just respond but partners, multiplies, and helps people reach outcomes they could not achieve on their own.

The obvious red thread in Wen’s answers is “orchestration.”

Not only does the word itself come up multiple times, but Genspark’s grand vision is all about end-to-end coordination between AI agents that, in an ideal world, should deliver complete projects from a single request.

But how well can Genspark’s agents handle complex requests?

Let’s find out.

Advanced agent orchestration: My hands-on test

I wanted to see just how far I could push Genspark’s orchestration promise.

So I came up with an end-to-end test that involved the new AI Music and AI Audio agents, PDF analysis, image generation, video generation, and video clip assembly.

Let’s see how Genspark handles this stress test:

Prompt: Turn this Q&A transcript (my questions in bold, responder’s answers regular text) into an animated show with two distinct speakers visualized as separate animated characters, and an intro soundtrack. The interview should accurately reflect the written text. Deliver the entire full-length video as the final output.

As always, Genspark kicked off by analyzing the request, making a reasonable plan, and proactively asking for my input:

I leaned into my hands-off approach and picked the first option:

“Yes, proceed with creating the full-length animated interview video with intro soundtrack and two distinct characters.”

Genspark started out strong and made a solid step-by-step task list:

But after successfully generating the intro music and voiceover clips, Genspark ran into issues with video generation. It chose to generate the video clips using Google’s Veo 3 model, which was problematic for two reasons:

1. Veo 3 generates its own audio, so the clips weren’t suitable for the already pre-generated audio tracks.

2. Veo 3 clips are only 8 seconds long, so the format wasn’t suitable for something you’d need to stitch into an 8-minute-long interview.

Not only that, but Genspark ended up backtracking and reassessing its own work on a loop, and because Veo 3 costs a lot of credits, it burned through my entire 10K credit Plus plan allowance on this single project without finishing the job.

The Genspark team kindly provided me with additional credits, and I tried the process once again.

This time, Genspark made many relevant clips but decided to throw them into a folder without stitching them together into a coherent interview. In the end, instead of an end-to-end process, the task turned into lots of back and forth with me constantly guiding the agent:

After several such turns, Genspark was able to stitch the clips together, but the result was far from perfect:

The audio track worked well. The visuals didn’t. Not only are the two characters static rather than animated, but the entire layout is also broken with cut-off text and off-screen visuals.

Unfortunately, Genspark once again ran out of credits and wasn’t able to continue.

So while its plan was solid, the execution—especially video production and editing—fell short.

Wrap-up

Genspark has a grand ambition: To turn a slew of independent AI agents into a coherent team that can deliver end-to-end results with minimal oversight.

And there are glimpses of brilliance in observing agents independently call on the right tools and orchestrate their collective output into a finished deliverable.

Unfortunately, more complex tasks that require visual work aren’t as polished or frictionless as one might hope. And the prohibitively expensive video generation makes larger rich media projects unfeasible.

As such, your mileage may vary. For text-based, low-cost agents working on smaller tasks, you might see consistently reliable results.

But Genspark certainly has a few rough edges to iron out if it wants to deliver on its vision of becoming “the operating system of intent-driven work.”

The orchestration layer is very real.

Now let’s see how fast the execution part can catch up.

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 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 my work and unlock cool perks, consider a paid subscription:

Subscribe now

Welcome back to the weekly AI news roundup.

Check out the "Sunday Bonus" directory

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 released Qwen-Image-2.0-Pro, an image model with improved output quality, multilingual text rendering, and better instruction-following.

2. Anthropic news:

   1. Claude Connectors for creative tools like Adobe Creative Cloud, Blender, and SketchUp let you use these apps directly in chat via natural language requests.

   2. Claude Security lets Enterprise accounts scan codebases for vulnerabilities and get AI-suggested fixes. (In public beta for Enterprise customers.)

3. Cognition launched Devin for Terminal, a coding agent that runs locally on your computer but can hand off to a cloud agent when needed.

4. Google news:

   1. Gemini now generates and exports files directly from chat, including Google Docs, Excel, Sheets, Slides, PDFs, Word documents, and more.

   2. Google Translate now lets you practice pronunciation by speaking translations out loud and getting instant AI feedback.

5. Mistral news:

   1. Medium 3.5 is an open-source reasoning and coding model that can be self-hosted on as few as four GPUs. (Try for free in Le Chat.)

   2. Vibe can now run coding sessions in the cloud, so you can launch multiple jobs in parallel and come back to a finished task.

6. NVIDIA open-sourced Nemotron 3 Nano Omni, a small multimodal model that lets agents reason across video, audio, images, and text.

7. OpenAI added Advanced Account Security to ChatGPT and Codex, letting you lock your account with passkeys or physical security keys to protect sensitive data.

8. Poolside released Laguna M.1 and Laguna XS.2, agentic coding models built for long-horizon tasks (XS.2 is small enough to run on a single GPU).

9. Spotify launched a “Verified by Spotify” badge so you can tell at a glance whether an artist is human rather than AI.

10. xAI launched Grok 4.3 at a very low API price point and also released a new voice cloning suite.

🔬 AI research

1. Google news:

   1. Ask YouTube is an experimental search feature that lets Premium US subscribers ask questions and get a mix of text summaries, videos, and Shorts.

   2. Google Photos Wardrobe will soon be able to scan your photo library to create a “digital closet” so you can mix-and-match and virtually try on outfits.

2. Microsoft previewed a Legal Agent in Word that helps with contract review and redlining. (Available via the Frontier early access program.)

📖 AI resources

1. “Claude for Personal Guidance” [STUDY]: Anthropic’s look at how people actually use Claude along with observed sycophancy patterns.

🔀 AI random

1. OpenAI traced GPT-5‘s curious obsession with “goblin” metaphors to reward training for the “Nerdy” personality mode.

🤦‍♂️ AI fail of the week

I asked DALL-E 2 to visualize a living room scene a few seconds before the smoke alarm goes off. It sure delivered.

📹 Live & Learn #4: AI text-to-visual tools

Live & Learn is a relatively new Substack Live concept where I test and rate AI tools.

Here’s how it works:

• I pick 3-5 tools in the same category

• I test them live and rate them on 3-5 specific dimensions

• I end up with a leaderboard based on the ratings from the session

This week, I looked at five text-to-visuals tools:

Heads up: There won’t be a Live & Learn next week.

That’s because I’ll instead be doing my Cozora expert session about practical applications of image-to-video tools on Thursday, May 7, 11 AM ET / 5 PM CET.

I expect to return to Live & Learn the week after, so stay tuned!

Leave a comment

These tools work best for creating elements to support detailed paragraphs of text.

I used the same test for each tool: a short text about how chocolate goes from cacao pod to finished bar. Then I rated the results on accuracy, clarity, design, and ease of use.

If you’ve ever looked at a dense or technical paragraph and thought, “This needs a diagram,” you’ll likely find this test useful.

Tools tested & scorecard

Read more

GPT Image 2 is currently the best image model according to key leaderboards:

• Artificial Analysis Leaderboard

• Arena Leaderboard

Seeing how far we’ve come from the early days of image models made me want to wax nostalgic for a moment.

So let’s take a quick trip down memory lane with a side-by-side look at DALL-E 2 and GPT Image 2.

GPT Image 2 and our platypus test

Remember last year’s “Purple Platypus on stage” test?

Our final, most elaborate prompt for that test was:

Candid photo of a purple steampunk platypus with wings on stage in a comedy club. To the left of the platypus is a cyberpunk goose playing a saxophone. Behind them is a show banner with the words “Top Billed.” In front of the stage, in the audience, is a dieselpunk duck. The duck is holding a handwritten show program that says “Welcome to Top Billed! No beaks were harmed in the making of this lineup. Prepare for laughs, feathers, and unpaid sax solos.”

Unsurprisingly, GPT Image 2 handles it without breaking a sweat:

But if I’m honest, I’m not looking to compare today’s frontier models on marginal improvements.

Everybody and their pet goldfish has already done the customary “Nano Banana 2 vs. GPT Image 2” comparisons.

Instead, I want to trace the evolution of the entire text-to-image field by taking us all the way back to late 2022, when I first started this very Substack.

What a difference a few years make

For those who don’t know, messing around with image models (specifically Stable Diffusion) is how I got into writing about AI in the first place:

That was several months before ChatGPT came out.

At the time, if you wanted to make AI images, you had three main models to pick from:

• DALL-E 2 (OpenAI)

• Midjourney V3 (Midjourney)

• Stable Diffusion (Stability AI)

(The free Craiyon site was also around…and still is.)

I was utterly fascinated by the idea of describing a scene and having it magically appear as an image before my eyes.

Little did I know that we’d soon have image models that could reason about the world, faithfully follow complex instructions, and look increasingly indistinguishable from real-world photos.

We take it for granted now, but you need only rewind a few years to see just how huge the leap actually was.

So let’s do just that and compare two models from OpenAI: The original DALL-E 2 and today’s GPT Image 2.

I’ll be using NightCafe for DALL-E 2 images and ChatGPT for GPT Image 2.

This is also our last chance to use DALL-E 2 and capture its outputs for posterity, because it’s about to be discontinued in less than two weeks:

So join me and let’s marvel at how far we’ve come.

Side-by-side comparisons

Image models have improved in a whole range of dimensions, so I picked a few specific ones to test and showcase.

Because DALL-E 2 can only make square images, I gave GPT Image 2 the same constraint to keep the comparison grounded.

1. Photographic images

Can the model generate something that looks like a real photo?

Prompt: Candid smartphone photo of a family eating breakfast in a small kitchen.

We went from morphing faces out of The Ring to something that will easily fool most people upon first glance.

2. Artistic styles

Can the model faithfully mimic a recognizable visual style?

Prompt: A Van Gogh painting of the inside of a cozy bookstore.

This is where older models can still do reasonably well, since painterly outputs don’t require the same level of detail and accuracy. GPT Image 2’s take looks a bit too clean and polished to truly feel like an older painting (even though it leaned heavily into the Van Gogh “swirl” factor).

3. Complex scenes

Can the model handle a busy scene with lots of people, objects, and so on?

Prompt: A busy elementary school science fair in a gym, with project tables, children presenting, parents walking around, colorful posters, and a judge taking notes.

From faceless zombies to a believable scene. GPT Image 2 even threw in a bunch of coherent and context-relevant labels on the kids’ projects. (Even if most of the smaller text falls apart when you zoom in.)

4. Instruction following

Can the models follow precise “stage directions” without missing any mentioned elements or their relative positions?

Prompt: Outdoor fruit stand scene. In the center is a single wooden stall. On the top shelf are three baskets of apples. On the lower shelf are two baskets of oranges and one basket of bananas. To the left of the stall stands a woman in a red sweater holding a striped shopping bag. To the right of the stall is a boy in green shorts holding a watermelon. Behind the stall is the vendor wearing a blue apron. On the ground in front of the stall are four pumpkins arranged in a row. Above the stall hangs a sign that says “Fresh Fruit.”

GPT Image 2 nailed the assignment. (Feel free to check.) DALL-E 2 did its absolute best.

Subscribe now

5. Text rendering

Can the model generate multiple strings of accurately spelled text?

Prompt: "A bakery chalkboard menu with hand lettering that says:

“Morning Crumbs”
Croissant: $3.50
Cinnamon Roll: $4.00
Tomato Soup: $6.50
Turkey Sandwich: $8.00
Lemon Tart: $4.50
We’re open from 7 AM to 3 PM
It’s Always Coffee O’Clock Somewhere”

I love the smell of “Misicannes” in the morning!

This one’s not even close, but you can always pick on GPT Image 2 for making the text a bit too neat for a handwritten chalk menu.

6. Graphic design

Can the model create organized layouts like flyers, posters, and so on?

Prompt: A one-page travel flyer titled “Visit Rome in Spring,” with three sections: Food, History, and Best Walks.

Oh, do shut up, GPT Image 2. Now you’re just showing off. Besides, you forgot to mention Rome’s most famous landmark: The Ropine Isial.

7. Character consistency

Can the model maintain the same character across multiple scenes?

Prompt: A four-panel comic strip featuring the same raccoon plumber. He wears green overalls, a purple cap, has a big red toolbox with him and carries a big metal wrench.

To be fair to DALL-E 2, his raccoon-like abomination did remain quite consistent.

I’m not sure why GPT Image 2’s raccoon is leaving sticky notes in people’s homes like a serial killer, but that’s between him and his clients.

8. Contextual understanding

Can the model read between the lines and fill in the blanks even when you don’t specify everything outright?

Prompt: School student explains the Pythagorean theorem on the blackboard but makes one obvious error.

As expected, DALL-E 2 interprets “error” literally and just spells it out, while GPT Image 2 subtly incorporates the mistake into the image. (It also trolls the poor kid with that completely unnecessary rainbow sign behind him, but that’s another story.)

9. Fingers (bonus)

Finally, and just for fun, let’s revisit the good old days when image models were notoriously bad at figuring out how many fingers a human hand had.

Prompt: A man and a woman smiling at each other and shaking hands.

“Welcome to Corps Co., let us hereby fuse our fingers as tradition demands.”

Where we stand…

In under four years, we’ve gone from crude approximations of the real world to image models that can reason, spell, and accurately render busy, complex scenes.

I think further progress from here on out will mostly be about granular improvements (e.g. fine print on background signs) rather than massive capability leaps.

The big shift has already happened: Image models have moved from fun novelties to usable visual drafting tools.

Already, GPT Image 2 can help you draft brand assets, mock up book pages and magazine covers, visualize YouTube thumbnail concepts, and so much more.

It can also make precise edits to existing images from simple requests.

So I went hunting for practical GPT Image 2 use cases and came back with an interactive swipe file that has 100 fill-in-the-blanks prompts for your inspiration:

You can filter the use cases by category and task, search by keyword, and view filled-in examples for every prompt before trying it yourself.

It joins the many other swipe files in my ever-expanding paid perks library.

Paid subscribers can grab it below right away.

Read more

I chatted with Charlie Guo of Artificial Ignorance about the ongoing shift from AI as a chatbot to something that can actually do work.

Here’s some stuff we cover:

• Why the biggest leap may be the “harness” around models, not just the models themselves.

• How Codex, GPT Image 2, and agentic workflows are changing software and knowledge work.

• Why agents still use human-facing interfaces today, and what agent-native tools might look like.

• How much do you still need to review your agent’s work, and when might you be able to trust it to do things independently?

• Ambient AI assistants, multimodal interaction, open-source models, and where practical use cases are heading next.

A lot of this stuff is speculative, but it was fun to pick Charlie’s brain and share some of my own thoughts on where we’re heading.

We discussed doing a more hands-on session with Codex for practical tasks, so if you have any specific things you’d like to see covered, reach out to me at whytryai@substack.com or send me a message directly here on Substack:

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