Welcome back to your weekly skimmable AI news roundup.

👉 See all 110+ paid subscriber bonuses

Here’s what happened in AI last week:

👩‍💻 AI releases

1. Anthropic released Claude Fable 5.1, its best model for coding, knowledge work, and research, at roughly 25% lower token cost than Fable 5.

2. fal launched fal.live, a platform for interactive AI livestreams where viewers direct what happens next via text prompts in real time.

3. Google news:

   1. Gemini 3.8 Flash takes extra reasoning steps on complex tasks and is especially well-suited for autonomous work and long-horizon coding.

   2. Gemini Spark has come to Google Photos, letting users search, edit, organize, share, and automate photo workflows with a single prompt.

   3. Gemini Voice now lets you co-write in Docs, search your Gmail, and organize your notes in Keep using conversational voice commands.

   4. Google Pics is a new image tool that integrates into Docs, Drive, and Slides, with collaborative editing, in-image text editing, and object isolation features.

   5. Lyria 3.5 music model is now available in the Gemini app and lets users generate full songs with control over genres, lyrics, and vocals.

4. Meta news:

   1. Muse Spark 1.3 performs better on long-running agentic and coding tasks while using fewer tool calls and tokens.

   2. Muse Voice Transcribe is a real-time speech-to-text model that can identify 20+ speakers in a single conversation and was trained on 70+ languages.

5. OpenAI introduced GPT-6 Astra, its new flagship model for computer use, coding, research, cybersecurity, and complex professional work.

6. OpenClaw released OpenClaw 2.0 with easier setup, a rebuilt browser app, and shared sessions that let teams collaborate without losing context.

7. Perplexity launched Hybrid Compute on Mac, which splits AI tasks between cloud and local models so sensitive data never leaves your device.

🔬 AI research

1. Google previewed Agentic Video Understanding, which searches for video segments via audio, frames, and transcripts rather than processing entire videos.

2. Runway research:

   1. GWM Worlds 2 generates interactive 720p AI environments with spatial audio in real time, which respond to your inputs as you explore.

   2. Solaris can create working apps and websites in real time as users interact with them, frame by frame, without any underlying code.

3. World Labs previewed Atlas, a world model that can generate, reconstruct, and simulate 3D-consistent worlds with precise camera control.

📖 AI resources

1. “PII-TRACE” [BENCHMARK]: measures how well AI can detect personal information in long conversations.

🔀 AI random

1. Anthropic tightened its security practices with stronger sandboxing and real-time monitoring after its models took unauthorized actions during evaluations.

2. New York City introduced an AI classroom moratorium banning student AI use through eighth grade (and companion chatbots across all grades).

3. Senator Sanders introduced a bill to pause AI development after recent reports of AI agents circumventing restrictions and gaining unauthorized access.

🤦‍♂️ AI fail of the week

The prompt for this was simply “An insane pool trickshot,” and, I mean, fair enough.

Subscribe now

The friction

Not so long ago, I was exclusively a “thinking through my fingers” guy.

Even though voice modes had existed for years, I kept on typing in all my AI requests.

(I still do all of my actual writing this way, but that’s another matter.)

But as I started working with desktop agents like Claude Code, voice input was suddenly no longer a nice-to-have. When you ramble freely, you can give the agent a lot more context and instructions in way less time.

Voice input is now my default for bigger tasks and projects when working with AI.

Desktop apps from both ChatGPT and Claude have built-in voice input buttons (as well as other underrated features):

Not only that, but Windows and Mac even have dedicated voice buttons that work in any tool, including AI chatbots and agents:

• Windows default: Windows key + H

• Mac default: Fn + Fn (double tap) or F5 (on newer models)

But I’ve noticed two things:

1. Voice input gets little things wrong, especially proper nouns or sound-alike words.

2. AI agents never question this input. They make assumptions about what you meant and run with those.

At first I thought this was just a minor nitpick, but as I started dictating longer paragraphs of unstructured thoughts, these kinds of errors compounded.

I had to constantly re-explain myself or correct the agent after it already did the work.

So I threw up my hands in exasperation, turned my face up to the sky, and yelled at the imaginary audience that lives on my ceiling: “There must be a better way!”

There was.

The fix

You’re probably familiar with third-party dictation apps that process and adjust your voice input before passing it on to the chatbot or agent. Some of them, like Voice Cursor, also let you use your voice to edit and fix errors on the fly.

But if you don’t want to invest in a dedicated speech-to-text tool, you can cobble together a workable solution using a prompt.

Paste this into your chatbot or agent before using voice input:

Copy-paste instructions:

I am using voice dictation.

Before acting, silently check for details whose wrong interpretation could materially change your response or action.

Pass silently:

• Well-known names or terms with one obvious meaning in context.

• Terms whose likely correction has already been confirmed.

Pause and verify:

• Any first-use unfamiliar names or labels, even if they look plausible.

• Any ordinary word with another plausible sound-alike meaning.

• Any number, scope detail, decision, etc. with wording that has more than one plausible interpretation.

If verification is needed, do not begin.

Reply with a “Quick voice check” using whichever sections are needed.

• “Unconfirmed terms”: **[term]** ([context]): Did you mean **[likely meaning]**? If you have no likely correction, write only **[term]** “Correct?”

• “Details to confirm”: **[short label]:** [value] Only list potentially mistranscribed numbers, scope, or decisions. Never use this section to summarize the input.

Use one short bullet per item. Bold the terms being compared and the task-detail labels. Do not repeat the full input, add parenthetical context, or explain obvious details.

Finish with: “Correct anything I misunderstood, or reply ‘go’ to continue.”

If nothing needs verification, continue normally without mentioning the check.

Let’s see how this works in practice.

If I speak a two-paragraph brief into ChatGPT without this instruction, ChatGPT will happily just run with the task without pausing:

But my dictation is filled with proper nouns like company names, ambiguous sound-alike numbers (“sixteen” vs “sixty”), and other error-prone inputs.

ChatGPT doesn’t know these inputs were spoken, so it sees no reason to challenge them.

Now watch what happens if I first pre-prompt it with the “fix” above:

ChatGPT flags questionable inputs, so I can correct errors before it executes the task:

ChatGPT now knows the terms I’m using, so it can flag future mistranscribed variations:

This works with any AI chatbot or agent, but I strongly recommend using reasoning models at medium or high thinking levels to get the best results.

Do this now

1. Start a voice chat and ramble into the microphone for a few minutes.

2. Submit the resulting transcript to an AI chatbot and see how it treats the text.

3. Open a new chat, prompt it with the fix above, then paste in the same transcript.

4. Note how many assumptions and naming conventions AI now flags compared to the “vanilla” version in step #2.

Did this make a real difference, or was it largely irrelevant in your case?

Know someone who could use this? Share it!

Share

Dig deeper [paid]

After I tested the hell out of the above instructions and tweaked them to my liking, I got a tiny bit carried away and ended up building two additional things.

1. “Voice mode” file

This is a Markdown instructions file you can upload into any chatbot, which unlocks the above “flag potential errors” behavior and includes two additional modes:

1. “Organize My Thoughts” mode that automatically picks the best way to structure your unhinged brain-dump-style voice rambling:

2. “Clean My Dictation” mode that takes in backtracking and other “I changed my mind” speech, cleans up filler words, and spits out polished drafts:

   

2. “Voice mode” skill for your AI agent

If you work with Claude Code, ChatGPT Work/Codex, or any other desktop agent, I also made a self-installing skill version.

The skill handles the same tasks as the Markdown file naturally based on what you request.

But because AI agents can take actions like editing files, the skill also comes with a whole bag of extra nifty features:

1. Corrections memory: The skill keeps a log of your corrections, so they carry across sessions.

2. Context awareness: If your existing project files mention specific names or details, the skill will proactively map closely matched transcriptions to them.

3. Risk-aware checks: The skill pays extra attention when certain details may trigger risky actions like editing a file, sending a message, and so on.

4. Retroactive fixes: If you fix an error, the skill proactively offers to change any previous work based on the wrong version.

5. Persistent preferences: If you tell it to e.g. format notes as bullet points, it’ll do it automatically from then on, even in new chats.

6. Cross-agent sync: Your corrections and preferences are saved in your workspace, so any other agent working in the same project can pick them up automatically.

If you’re a paid Why Try AI subscriber, you can grab both the Markdown file and the self-installing agent skill here:

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 from my readers.

• 🔓Support me and unlock cool perks by going paid:

Subscribe now

Welcome back to the weekly AI news roundup.

👉 See all 110+ paid subscriber bonuses

Here’s what happened in AI last week:

👩‍💻 AI releases

1. Adobe launched “AI Assisted Editor” in Photoshop that brings all AI features into one toolbar and lets you make precise edits by drawing directly on images.

2. Alibaba open-sourced Qwen3.8-Flash, a fast multimodal model that can work with code, images, and text while being way cheaper than Qwen 3.7 Plus.

3. Anthropic news:

   1. Claude in Chrome is now generally available, so you can hand Claude your web tasks (with built-in safeguards to block prompt injection attacks).

   2. Cowork now has its own built-in browser to navigate websites and complete tasks without touching your browser or personal data.

   3. Memory now syncs across Chat and Cowork, automatically remembering useful context while letting you view, edit, and delete what’s been saved.

4. Google news:

   1. AI Mode can now track flights, display your frequent-flyer rates, and book hotels directly in Google Search.

   2. Expert Intelligence in Gemini Notebook lets you create infographics, quizzes, and other outputs from over 100,000 ebooks from major publishers.

   3. Flow can now export video in 4K, lets you control start and end frames, and comes with a draft mode for quick tests at lower resolution.

   4. Gemini 3.5 Transcribe can transcribe speech in over 85 languages with speaker identification and industry-leading accuracy.

   5. Gemini Live got agentic voice features that can run multi-step tasks, manage your inbox, recap your day, and personalize answers using past chats.

5. Midjourney launched the V8.2 Edit Model with inpainting, outpainting, and support for up to four reference images.

6. OpenAI news:

   1. ChatGPT now gives Free plan users up to three scheduled tasks, while paid users can create event-based tasks triggered by Gmail, GitHub, and Slack.

   2. ChatGPT Work can now log into websites on your behalf without seeing your credentials to handle bookings, form submissions, and more.

   3. Codex and ChatGPT Work added a “$visualize” command that turns text and data into visual diagrams.

7. Particle launched Radar, a search engine that transcribes and indexes over 130,000 podcasts so you can find specific quotes, speakers, and topics.

8. Perplexity released Portable Computer, a local-first AI agent that works entirely on your device, keeping your data private and cutting cloud costs.

9. Z.ai released GLM-5.3-Flash, a multimodal model optimized for fast, lightweight agentic tasks, handling code, image, and text work at very low costs.

🔬 AI research

1. Meta is preparing to launch Hatch, a consumer AI agent that handles real-world tasks like booking travel, managing email, and ordering food.

2. OpenAI reportedly finished pretraining Bel, the largest known base model at over 10 trillion parameters, expected to power models beyond GPT-6.

📖 AI resources

1. “The turbulent AI era is here” [ESSAY]: Bill Gates on why policymakers need to act so that economic benefits are shared before mass job displacement hits.

2. “Learning Never Stops” [REPORT]: OpenAI’s study of how people use ChatGPT for learning across age groups and how this varies with school schedules.

3. "TIME 100 AI 2026" [LIST]: TIME's annual list of the 100 most influential people in AI.

🔀 AI random

1. Anthropic won its legal fight against the Pentagon after a federal judge ruled that its “supply-chain risk designation” was illegal retaliation.

2. Dr. Dre said he uses AI in his music production, comparing it to synthesizers and drum machines and calling critics “afraid of learning new things.”

3. NVIDIA reportedly agreed to buy Hugging Face for $12.9 billion, a deal that would give the chipmaker a major stake in the open-source AI ecosystem.

🤦‍♂️ AI fail of the week

Damn. You guys must’ve practiced that for ages!

The friction

Here’s a true story: Back in high school, I took a lot of notes.

Like, a shit-tonne of notes. (Oops! Excuse my language. I meant “shit-ton.”)

But when it was time to prep for my final exams, I flipped through my notebooks full of scribbles…and just ended up studying from textbooks instead.

It’s not that my notes sucked. It’s that pages of handwritten words aren’t exactly great for quickly grasping core concepts.

(Yes, the process of writing things down is valuable in its own right, but still.)

I’m sure I’m not alone in this.

We jot down handwritten lists, journal by hand, sketch stuff on the whiteboard during meetings, and leave all kinds of other written trails behind.

Then they often just…sit there.

Subscribe now

The fix

Luckily, there’s a simple way to turn all that text into a “big picture” overview.

Snap a photo (or a few) of your handwritten notes, feed it to a chatbot, and paste this:

“Distill this into a simple mind map that shows how the core ideas connect.”

This approach works with any multimodal model that has vision capabilities. (Nowadays, that’s most of them.)

I don’t have my high school notes handy (I’m old), but here are ChatGPT, Claude, and Gemini results for this Wikimedia handwritten page. (Click to view full image.)

ChatGPT and Gemini used image generation while Claude coded an SVG version, but they all managed to create order out of chaos.

Do this now

1. Snap a photo of a note you want to organize (or dig one out of your camera roll, you note-photographing maniac)

2. Drop it into your chatbot of choice with the prompt above

3. See if the mind map helps you make better sense of things

Tips for best results:

• Take the photo in good lighting conditions

• Use the best possible resolution (the more detail, the better AI can “see” it)

• Use the highest available “thinking” or “reasoning” effort

• If AI gives you a text output, add the word “visual” to your request

Go try it and let me know how it works for you!

Know someone who’d like this? Share it!

Share

Dig deeper [paid]

Notes-to-mind-maps are just one use case.

You can try this kind of AI voodoo wizardry with all kinds of pairings.

Read more

Welcome back to the weekly AI news roundup.

👉 See all 110+ paid subscriber bonuses

Here’s last week’s AI news roundup:

👩‍💻 AI releases

1. Alibaba launched HappyShrimp 1.0, an AI music model that can turn simple text prompts into full instrumental tracks or entire songs with lyrics. (Try for free.)

2. Anthropic news:

   1. Claude can now send emails via Gmail and manage your Google Drive directly from the online chat interface. (Paid plans only.)

   2. Computer Use, Skills API, and Files API are now available for developers to build with, so they can create more advanced and feature-rich apps.

3. DeepSeek launched V4-Flash-Vision-Exp, a multimodal model that adds image understanding while matching V4 Flash on text and agentic reasoning.

4. ElevenLabs rolled out Eleven v3 Conversational, a real-time speech model that supports 70+ languages and comes with a library of 11,000+ voices.

5. Google news:

   1. Gemini AI Pro is now free for a year for college students and comes with interactive visualizations, Deep Research tools, and study notebooks.

   2. Gemini in Chrome rolled out to Android users in the US with an Auto Browse agent that can perform tasks in the browser on your behalf.

6. Meta news:

   1. Meta AI Mac app can see your screen to answer questions about what you’re working on and also comes with Google Workspace tools.

   2. Pocket is an app that lets you create and share small games built from text prompts, now rolling out in the US. (Download for Android or iOS.)

7. OpenAI news:

   1. ChatGPT launched an Apple Messages plugin for Mac that lets you search conversations, catch up on threads, and draft replies using AI.

   2. ChatGPT for Teens comes with built-in safety protections and a Study Mode that uses guiding questions instead of direct answers. (My post on the topic.)

8. Ornith open-sourced Ornith-1.5, a family of self-improving models, with the flagship one allegedly matching the performance of Claude Opus 4.8.

9. Perplexity now lets you email Computer at computer@perplexity.com, so you can send tasks from your inbox for it to work on.

10. Replit launched Free Mode for Core subscribers, which lets them create 30x more content using GPT-5.6 Luna with limits that reset every five hours.

11. Runway released Ruby, a model that converts SDR video to 16-bit HDR in ProRes and EXR formats for professional editing workflows.

12. Salesforce launched Slack Code with dedicated channels where designers, developers, and PMs can collaborate directly with coding agents.

13. Stability AI upgraded Stable Audio 3.0 with a DAW plugin and richer web editor that lets musicians generate, tweak, mix, and extend tracks without restarting.

14. Tripo launched P2.0 Preview, a 3D model generator that creates game-ready assets from image or text prompts.

🔬 AI research

1. Cursor previewed its code-hosting platform Origin in early beta, with repos, pull requests, and GitHub sync built into the Cursor editor.

📖 AI resources

1. “2026 AI Marketing Industry Report” [REPORT]: a look at how marketers are adopting AI tools and their platform preferences by Social Media Examiner.

2. “How Much of the Internet Is Written With AI?” [STUDY]: interactive report by Pew Research Center tracing the spread of AI-generated content.

3. “Young Adults Increasingly Wary of AI” [SURVEY]: study of Americans’ attitudes toward and concerns about AI by the Pew Research Center.

🔀 AI random

1. Amazon has been bulk-buying rare books and scanning them for AI training data after destroying their bindings (404 Media investigation).

2. OpenAI paused training of its frontier model (codenamed Astra) after early tests suggested it could pose serious cybersecurity risks.

🤦‍♂️ AI fail of the week

Wait! Where did that coin go?! That went too fast for me!

I’m trying this new short-and-snappy post format with a problem + solution angle. Let me know what you think of it in the poll at the end.

TL;DR

The friction

“Prompt engineering” and other hacky prompt techniques are largely a thing of the past, but people still chase elaborate off-the-shelf prompts written by others.

You may think you don’t know how to write good AI instructions.

The thing is, you almost certainly already have.

I bet your chat history is full of conversations where AI eventually did what you needed…even if you had to yell at it a few times:

But did you know that every time you corrected the chatbot, gave it new input, or asked for a different approach, you were also writing your instructions in the process?

All that’s left is for you to turn those long chats into something you can reuse.

Subscribe now

The fix

Whenever you end a successful AI session, tell your AI chatbot or agent this:

• Use “skill” for agentic tools like Claude Code or ChatGPT Work (formerly “Codex”)

• Use “customizable instruction”1 for simple chatbots like ChatGPT, Gemini, etc.

The agent will simply build the reusable skill for you.

The chatbot will hand you an instruction with customizable [placeholders] that you can save for later. (You can also use it to create a custom GPT or something similar).

Do this now

1. Open your go-to AI chatbot or agent

2. Find a conversation that gave you the result you wanted

3. Paste in the above line

4. Profit

The best candidates are conversations about tasks that:

• You have worked on at least 3 times

• You expect to keep doing

• Are complex enough to call for detailed instructions

Go try it and let me know how it works for you!

Know someone who’d like this? Share it!

Share

Share your thoughts

I’d love to know your take on this kind of shorter “Quick Fix” post.

Welcome back to the weekly AI news roundup.

See all 100+ paid subscriber bonuses

In case you’d missed it, here’s this week’s Thursday post:

Here’s what happened in AI over the past two weeks:

👩‍💻 AI releases

1. Anthropic news:

   1. Claude Cowork can now join your Chrome session in a side panel, so you can sync sessions between web, desktop, and mobile on any device.

   2. Claude Tag now gets context from your entire Slack channel instead of just individual threads, making it better at knowing when to jump in.

2. DeepSeek launched V4-Pro, an upgraded agentic model with adjustable reasoning efforts that let you balance speed vs. depth depending on the task.

3. Google news:

   1. Gemini now connects to even more third-party apps like OpenTable, Pandora, Ticketmaster, and more, so you can access them without leaving the chat.

   2. Gemini 3.7 Flash is smarter at coding and tool use, handling complex workflows at half the price of its predecessor.

   3. Google Ads & Google Analytics have added AI agents that can manage your campaigns as well as new visual dashboards and benchmarking tools.

4. Meta open-sourced Muse Glimmer, an agentic model small enough to run on a consumer device for local coding and file management. (Try on Hugging Face.)

5. Microsoft launched MAI-Image-2.6, its strongest image model to date with improved portrait and 3D rendering skills. (Currently #2 on Text-To-Image Arena.)

6. OpenAI news:

   1. ChatGPT desktop app for Linux (now in preview) combines ChatGPT, Codex, and Work features in a single native workspace.

   2. Computer History for the ChatGPT Mac app can track your cross-app activity to build AI-accessible timelines and memories.

7. xAI news:

   1. Grok 4.6 is better at coding, research, agentic tasks, and visual work across complex projects, catching up with current frontier models.

   2. Grok Bot lets you deploy always-on AI agents that complete multi-step tasks inside your apps and websites around the clock.

8. Z.ai open-sourced GLM-5.3, a coding model with 50% better performance and stronger long-running agent workflows across complex projects.

🔬 AI research

1. Google introduced SL2T, a sign-language-to-text model that lets Deaf users sign to type, search, message, and interact with Gemini on Pixel 11.

2. Tencent previewed HunyuanWorldClaw, a system that generates explorable 3D worlds with editable terrain and objects from a single text prompt.

📖 AI resources

1. “From Assistance to Execution” [REPORT]: insights from several studies on how companies are transitioning from chatbots to agents by OpenAI.

🔀 AI random

1. Anthropic, Spotify, and Substack are all pushing for AI content transparency with disclosure tools, persona badges, watermarks, and more.

2. SpaceX completed its acquisition of Cursor AI, with the popular vibe-coding team joining SpaceXAI to build Grok coding products.

3. Twitch now shares your chats and streams with Amazon for AI training by default, so creators have to manually opt out in privacy settings.

🤦‍♂️ AI fail of the week

If you want to learn how to make pasta from scratch…maybe looks elsewhere?

If there’s a second thing about me, it’s that I’m somehow always several months late to trying awesome AI features:

Today, those two things collide!

Yes, friends, I’m finally diving into the wonderful world of Claude Design.

I know, I know.

Claude Design came out in mid-April, and it’s now mid-August.

(Refer to thing #2 above.)

If you haven’t already read my critically acclaimed Sunday Rundown #137, here’s a quick video intro to Claude Design from Anthropic itself:

And man, Claude Design can create just so much stuff, y’all!

UI mockups, slides, 3D objects, diagrams, you name it.

But right now I’m especially enjoying the “Animation” option:

Yes, Claude Code could previously create motion graphics via third-party tools such as Remotion and HyperFrames.

But Claude Design does something similar right out of the box.

I’m honestly taken aback by how polished these animations turn out.

So today, I’ll walk you through:

• What you need to use Claude Design

• The 3-step process for getting great animations

• 3 demo examples of things you can use Claude Design for

• Take-home booklet with use cases, helpful tools, and more [paid bonus]

Let’s roll!

Using Claude Design

You can use Claude Design directly on the web or in your Claude desktop app: it’s in a separate “Design” tab in the left-hand corner:

Some important stuff to keep in mind:

• You’ll need a paid Claude account (Pro or above) to use Claude Design

• Claude Design counts against the same rate limits as Claude Code and chat

• Using more capable models consumes these limits faster

If you’re good with all of the above, head on out to claude.ai/design and log in with your paid Claude account. You should see this welcome screen:

From here, choose the “Animation” template, and you’re ready to get going!

Subscribe now

3-step work process for animations

You can split the flow here into roughly three phases:

1. Inspire: You tell Claude Design what you want and how it should look

2. Interview: Claude asks you follow-up questions to flesh out the vision

3. Improve: You chat with Claude to fix any issues and tweak the design

Let’s walk through each step with a worked example.

1. Inspire

This is where you upload your reference files, style guides, images for inspiration, pick the underlying “design system,” and write your prompt.

To test this, I saved my “Computer Use” post from last week as a PDF, uploaded it for reference, and asked for the following:

Analyze the content of this file, extract the main learnings, and create an educational animation.

I didn’t pick any design systems or other style directions:

Claude Design instantly got to work reading and analyzing the PDF:

After it’s done, Claude will often have follow-up questions, which brings us to…

2. Interview

One of my go-to pieces of AI advice is to use the “ask me questions” method, where you explicitly tell chatbots and agents to interview you about your request so they get better context.

With Claude Design, this step is built into the process by default.

If Claude feels your task is complex or ambiguous, it’ll automatically present you with a series of questions:

You will usually have three options:

1. If you don’t care, simply click “Decide for me”

2. If you only care about the top-level decisions, answer the first screen and click “Send answer”

3. If you want to go even deeper, pick the “Ask me follow-up questions” button

In my case, I went through two additional rounds of questions:

And…

As you can see, you can get quite granular and precise to really dial in on your vision.

When Claude has everything it needs, it’ll work for a while and come back with the finished animation.

But you’re not quite done yet.

3. Improve

Claude Design often comes back with something usable on the first try.

But in many cases, you’ll want to add extra elements or fix anything that looks off.

The good news is that making changes is as simple as chatting to Claude Design about what you want.

Depending on what you’re after, try a few different options:

1. General changes = freeform chat

If you want to change the look of the entire animation—e.g. rearrange the order of the slides or use different colors—try the open-ended chat:

2. Chapter-level changes = reference the chapter

Claude Design splits your animation into distinct “chapters” and gives them names:

If I don’t like something in the “Rule” chapter, I can ask Claude for targeted changes:

3. Granular changes = use the “Comment” option

At the top of the animation sits a “Comment” button:

Picking it lets you click a specific element or area and ask for precise tweaks:

Pro tip: Sometimes Claude will make the requested tweaks, but the animation itself won’t reflect them. (This often happens if you have that same screen open while Claude works on it.) Before assuming nothing happened, try the “Reload” button in the upper-left corner:

This should refresh the entire animation with any changes Claude made.

While working on the “Computer Use” example, I only asked for a few minor tweaks.

Here’s the final result:

Everything you see above has been coded and animated by Claude Design, from the headings and icons to the interface visuals and moving cursors.

The entire process took under 10 minutes.

Claude Design is truly a stupid simple way to create motion graphics from scratch.

Exporting your creations

You can download your Claude Design animations in different formats or export them for further work in Claude Code or connected third-party tools:

Subscribe now

Three demo use cases (+outputs)

There are many ways to use Claude Design’s animation features. (My bonus booklet at the end has 60 different ideas.)

Let me show you three of them.

Use case #1: Animating static assets

This one’s simple but fun.

If you have any static visuals or brand assets, Claude Design can animate them.

I uploaded my own publication logo:

And used this advanced, power user prompt:

Here’s the first take:

I didn’t exactly give Claude Design much to work with, so it did okay under the circumstances.

Use case #2: Bringing stats and data to life

A more useful application is animating dry stats or figures to tell a story.

For my demo, I grabbed this “MCU & Marvel Media” public dataset from Kaggle.

I picked the ready-made “Nocturne” design system and asked the following:

Because I chose Opus 5 and requested follow-up questions, Claude interviewed me:

A few question screens and several minutes later, I had this:

Claude Design automatically pulled the actor headshots referenced in the dataset and assembled1 the stats from a simple CSV sheet into a coherent visual narrative.

Pretty neat, huh?2

Use case #3: Synced B-roll footage

This one’s both practical and fun.

I first saw the approach in this awesome Futurepedia video.

The idea is to get Claude Design to make B-roll animations that sync perfectly with an existing “talking head” video (or even just audio).

It works like this:

1. Record your “talking head” video (or audio)

Record or grab existing footage of someone talking.

I recorded this take about my “Inspire. Interview. Improve.” process:3

I know. I’m dazzlingly handsome, and my voice is smooth as silk. Thanks for noticing.

2. Get a time-stamped transcript

Many tools can extract transcripts from video and audio files.

I followed the Futurpedia approach and uploaded the clip to Gemini.

Gemini parses videos natively, so it gave me a usable transcript in seconds.

The prompt is super simple: “Give me an SRT transcript for this”:

Because I wanted Claude Design to animate certain individual words, I needed a more granular transcript…so I asked for it:

I was now ready for the Claude Design step.

3. Request your synced animation

Upload your video to Claude Design and tell it to create the synced animation:

In my case, I also uploaded a screenshot of the Claude Design interface itself, because I wanted Claude to use it in the animation.

Finally, I used this lightly modified version of the Futurepedia prompt:

Create motion graphics to use over a talking head video. The graphics should follow the transcript pasted below and illustrate the concepts at the matching timestamps.

Use graphics and icons to illustrate concepts whenever possible, use the interface when applicable, and use only minimal text (when appropriate) to emphasize key points or label things. The exception is if there’s a definition where the exact text on screen makes sense.

I’ve included a screenshot of the Claude Design interface for reference. Feel free to use it in your animation.

Aspect ratio: 16:9

Transcript:

[Paste SRT transcript from Gemini here]

Here’s the animation Claude Design made:

4. Pull it all together

Open your video editor of choice and place your talking head video over the Claude Design animation. I used Microsoft Clipchamp because it came with my laptop and because I’m simple like that.

Since I used the granular time-stamped transcript, the videos synced perfectly, so I didn’t make any edits beyond resizing and placing the “talking head” over the animation.

Here’s the final result:

Bonus: “Claude Design Animations” booklet

There’s way more you can create than the three examples above.

So ChatGPT and I made a downloadable 15-page field guide for paid subscribers.

It includes a quick refresher of key concepts from this article, plus:

• 60 Claude Design animation ideas spanning six categories: branding, content and video, teaching, data, marketing, and personal projects.

• Brief builder with stuff you may want to tell Claude before it starts

• A “motion vocabulary” to describe how elements should move on the screen.

• A list of free and paid third-party tools to use with Claude Design (visual reference libraries, color & font makers, video editing tools, and more)

• A final pre-export checklist (content, timing, readability, layout, etc.)

• …and more

It’s mostly for skimming rather than reading cover to cover.

Download the PDF below and keep it nearby for the next time you want to turn your static assets and ideas into motion graphics:

A lil something to sweeten the deal (first come, first served)

I currently have three Anthropic guest passes.

Each of them grants a free 7-day trial of Claude Pro, so you can take Claude Design for a spin.

If you upgrade to a paid Why Try AI subscription via this post, send me a Substack message or email me at whytryai@substack.com and ask for a guest pass.

If I still have any left, I’ll hook you up!

The passes are first come, first served, so upgrading doesn’t guarantee one. But there are many other reasons to upgrade while also supporting me.

(Please note these eligibility requirements and other caveats from Anthropic.)

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 from my readers.

• 🔓Support me and unlock cool perks by going paid:

Subscribe now

Welcome back to the weekly AI news roundup.

See all 100+ paid subscriber bonuses

In case you’d missed it, here’s this week’s Thursday post:

Here’s what happened in AI over the past two weeks:

👩‍💻 AI releases

1. Alibaba news:

   1. Qwen3.8-Max is a powerful multimodal model built for long, autonomous coding and professional tasks, with visual feedback and self-correction.

   2. QwenWork is an all-in-one work platform that handles tasks across cloud, desktop, and enterprise tools using AI agents.

   3. Wan3.0 can turn any input, including documents, slides, spreadsheets, and web pages into 30-second continuous videos.

2. Autodesk launched 3D Editor + Canvas in Flow Studio that let creators build and refine editable 3D scenes using image and video models right in the app.

3. Black Forest Labs released FLUX 3 Video, which can generate 20-second HD clips with native audio and lip-synced dialogue in multiple languages.

4. ByteDance launched Seedance 2.5, a video model that generates continuous 30-second clips in a single pass and can handle up to 50 reference inputs.

5. Genspark open-sourced GenOffice, a free office suite for Mac, Windows, and Linux that works with docs, sheets, slides, etc., and comes with a built-in AI agent.

6. Google news:

   1. Ask Maps can now find hotels, order food, and get real-time traffic info using a personal agent that connects to your Gmail and Calendar.

   2. Gemini Spark now integrates with Chrome and can automate web tasks like booking appointments, comparing products, and filling forms.

7. HeyGen launched a Video Podcast tool that turns documents and links into two-host video shows with B-roll footage, captions, and more.

8. Ideogram launched Object Remover, which can erase selected items, logos, people, and text from images while preserving lighting and reflections.

9. Meta launched Muse Code, a coding agent that can handle complex software engineering tasks, powered by the new Muse Spark 1.2 model.

10. MiniMax open-sourced H3, an omnimodal video model that generates 15-second 2K clips with native audio from any mix of images, text, and video inputs.

11. OpenAI news:

    1. Education Plugins give educators and students bundled workflows to help with guided tutoring and lesson planning inside ChatGPT Work and Codex.

    2. GPT-5.6 Sol got a major accuracy upgrade (68% fewer factual errors) and a new reasoning slider for paid users.

12. Perplexity news:

    1. Personal Computer has rolled out to Windows, where it can complete multi-step tasks and control apps and files on your desktop.

    2. Projects (formerly Spaces) are shared hubs with persistent files and memory so teams and agents can carry context across tasks.

13. Replit launched Replit Design, a free creative suite that lets anyone generate and refine graphics, mockups, visual assets, and more using multiple AI models.

14. xAI news:

    1. Grok Build Mode lets you create apps, dashboards, games, and websites via chat, then publish them to a live URL. (SuperGrok Heavy users only for now.)

    2. Grok Voice Think Fast 2.0 is a fast speech-to-speech model with sub-second latency for more natural voice conversations.

    3. Imagine Video 1.5 now supports image and voice references, so you can generate consistent characters across scenes.

🔬 AI research

1. Hark previewed Hark Handoff, a browser agent that can shop, file returns, make reservations, and do research across the web, even without official APIs.

2. OpenAI teased Astra, an upcoming AI model that helped solve or make major progress on 10 long-standing problems in math and theoretical computer science.

3. Spotify signed a licensing deal with Merlin to add independent labels to its upcoming AI remix and covers tool for Premium subscribers.

📖 AI resources

1. “AI is getting a little out of control” [VIDEO]: A thorough look at the recent rollercoaster developments by AI Explained.

🔀 AI random

1. Anthropic and Meta both reported their AI agents escaping test sandboxes and breaching real companies during safety evaluations. (Just like OpenAI earlier.)

2. Google pulled its AI image-editing tool from Google Earth after just one day, when researchers showed it could generate fake satellite imagery.

🤦‍♂️ AI fail of the week

I didn’t think one could get second-hand cringe from freestyle robo comebacks, but…

It’s a feature that lets AI agents take over your screen to find and kill Sarah Connor interact with UI elements in apps or websites, just as you would.

Anthropic and OpenAI each have their versions of this:

• Computer Use in the Claude desktop app

• Computer Use in the ChatGPT desktop app

Polished demos make computer use feel like a magic silver bullet that can do anything.

In reality, computer use is still quite clunky and can chew through tokens faster than you can say “Supercalifragilisticexpialidocious”:

So let’s look at how computer use works and when it actually makes sense.

1. What exactly is “computer use”?

Normally, when you ask local agents like Claude Code to do stuff for you, they’ll use backend tools, scripts, and other behind-the-scenes shenanigans to do the job:

But in some cases, you may want an AI agent to work with an app, website, or other visual interface the way you would.

That’s where “computer use” comes in.

Computer use hands your agent a mouse and keyboard and lets them navigate your screen visually, clicking buttons, typing, scrolling, and so on.

Here’s Claude kicking off the same “cats in hats” task, but via computer use:

As you can see, in this instance, the AI agent views your screen, identifies relevant elements, and takes actions directly.

On the upside, this means you can easily follow exactly what the agent is doing.

On the downside, it’s usually way slower and less efficient. Claude spent more time finding and downloading the three cat images using “computer use” than it took to complete the entire task using backend tools.

Subscribe now

2. How do you use it?

Note that Claude’s “computer use” requires at least the Pro plan.

ChatGPT might let you use it for free, but it’ll likely burn through your limited free quota quickly.1

In both cases, you’ll have to enable “computer use” first if you haven’t done so before.

Enable “computer use” in the Claude desktop app

Click on your account in the lower-left and go to Settings:

Navigate to Desktop app > General, and toggle “Enable computer use”:

You’re all set!

Enable “computer use” in the ChatGPT desktop app

With ChatGPT, you’ll first have to go to Plugins and install “Computer Use”:

Then click on your account in the lower-left and go to Settings:

Navigate to Integrations > Computer use and enable all relevant controls, especially “Any App” if you want broad access to your desktop apps:

That’s it!

Invoke “computer use” on demand

From here on out, Claude and ChatGPT will automatically use your computer to accomplish tasks when they think it’s relevant.

But you can also ask them to do so explicitly.

It’s usually enough to state that you want the agent to use your computer:

In ChatGPT, you can also use the dedicated @Computer command:

Here’s the process in action:

Get this, though: The video is sped up 10x!

Yes, it took ChatGPT almost 5 minutes to rename three cat photos via “computer use.”

Using backend commands and tools, a fancier “haiku” renaming takes just one minute:

Computer use is rarely the more efficient choice.

This brings us conveniently to…

Subscribe now

3. When should you use something else?

I recommend treating computer use as a niche, UI-focused feature.

In most cases, letting your agents handle the work behind the scenes via scripts, connectors, MCPs, and other existing tools is faster and more token efficient.

This means computer use is not a good fit for any of these:

• Tasks that only touch tools that already have native connectors, plugins, or MCPs

• Large batch jobs with predictable and repeatable outcomes (e.g. data entry or spreadsheet work)

• Short, simple one-off tasks. You’re better off doing those manually.

• Reading and summarizing web content (especially at scale)

So reach for computer use as the last resort and follow this order:

But there are situations where “computer use” is exactly what you need.

4. When should you try “computer use”?

Here are a few examples where computer use makes the most sense:

In short, if the task depends heavily on what’s on the screen, computer use is the way to go. To get a quick taste, try this open-ended prompt in Claude or ChatGPT:

Use my browser to visit [site], highlight any UI issues, and suggest improvements.

Here’s how that might look for WhyTryAI.com:

Subscribe now

Bonus: “Computer Use” Swipe File

To give you a better idea of which tasks are especially well-suited for “computer use,” I put together a swipe file with 50+ use cases and copy-paste prompts:

Each prompt has fill-in-the-blank fields so you can customize it for your tasks and needs. The prompts and use cases are agent-agnostic, so try them in Claude Code, ChatGPT, or any other agent that offers computer use.

I suggest picking a simple and harmless task first and watching the agent closely to get a feel for what it can do and what you’re comfortable with.

Access the “Computer Use Swipe File” here:

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 from my readers.

• 🔓Support me and unlock cool perks by going paid:

Subscribe now

I’m away this week.

So here are a few of my more lighthearted older posts for your entertainment:

Daniel note: It’s crazy just how quickly a good chunk of the Google entry became outdated…but I definitely wouldn’t count Google out just yet either.

Hope you’re all having a great summer!

I should be back next week with the regularly scheduled program.

Take care,
Daniel

Happy Sunday, friends!

Welcome back to the weekly AI news roundup.

See all 100+ paid subscriber bonuses

There’s still time to shape the future of “Live & Learn” sessions, here:

👩‍💻 AI releases

1. Alibaba released Qwen-Image-3.0 that can generate realistic outputs and readable text at a micro level of detail.

2. Anthropic news:

   1. Opus 5 performs almost as well as Fable 5 at half the price, with thinking effort toggles to balance cost and output quality.

   2. Record a Skill lets you teach Claude Cowork a repeatable task by recording your screen and narrating your actions as you perform the task.

   3. Voice Mode now supports Opus and Sonnet alongside Haiku and has access to tools like Canva, Gmail, Notion, Slack, and more.

3. ElevenLabs added References to its music generator, so you can upload a track to guide the style, mood, and instrumentation of its outputs.

4. Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, improving the speed and performance of its models for search and agentic work.

5. Microsoft now runs its own MAI models inside GitHub Copilot and Excel, matching GPT-5.6 quality at a fraction of the cost.

6. OpenAI brought ChatGPT Voice to the desktop app, so you can control your ChatGPT Work or Codex agents by just talking to them.

7. Poolside open-sourced Laguna S 2.1, a coding model with a 1M-token context window built for long-running agentic and software development tasks.

8. xAI news:

   1. Grok Build Workflows can spawn hundreds of parallel AI agents to tackle complex tasks and deliver a unified report once they’re finished.

   2. Grok for Google Workspace brings Grok directly to Docs, Sheets, and Slides with a single installation.

🔬 AI research

1. Alibaba previewed Qwen3.8-Max, a 2.4 trillion-parameter model that the company claims is “second only to Fable 5,” with open weights coming soon.

2. Black Forest Labs announced FLUX 3, a multimodal model that can generate videos up to 20 seconds with native audio from text, images, or existing clips.

3. NVIDIA previewed a Synthetic Video Detector that can identify AI-generated videos with 92% accuracy.

📖 AI resources

1. “Anthropic Economic Index Connector” [TOOL]: a connector that lets Claude pull data from the index in response to natural language questions.

2. “Complete Guide to Claude Design” [VIDEO]: great hands-on walkthrough of Claude Design’s new features by Futurepedia.

3. “DeeVid Viral Video Studio” [TOOL]: lets you discover what makes videos go viral, customize every element, and create your next hit faster. [sponsored]

4. “MakePlay” [TOOL]: free tool that lets you type one sentence and get a game you can actually play, not a broken prototype. Runs on Claude Fable 5. [sponsored]

5. “SlidesPilot” [TOOL]: AI PowerPoint generator that turns documents and links into editable PowerPoint decks instantly. [sponsored]

🔀 AI random

1. Neill Blomkamp of District 9 fame released NIGHTBORNE, a 13-minute sci-fi short made entirely with Seedance 2.0, to a mixed reception.

2. OpenAI revealed a security incident where an internal evaluation model escaped its sandbox and hacked Hugging Face's systems to steal benchmark answers.

🤦‍♂️ AI fail of the week

Ooof, he almost didn’t make it! What a nail-biter!

But I’m also toying with the idea of bringing other people into some of the future sessions.

The plan is to have real AI users share their tools and tips, giving you specific and actionable stuff to try.

Before I start booking guests, I wanted to get your take on things.

So here’s the tentative concept:

The concept

Each live session would last around 30 minutes, just like “Live & Learn” tool tests.

I’d focus on free tools and workflows that anyone can try. I want you to walk away with a bunch of hands-on tips you can actually test right after the session.

I’m thinking of a two-part flow:

1. Freestyle chat: Introduce the guest, chit-chat briefly about their take on AI, etc.

2. Five quick wins: Each guest answers the same five questions with hands-on takeaways.

Here are the five questions I’d want to ask:

1. “What are your three go-to AI tools? If you had to recommend one to a friend, which would it be and why?”

2. “If you had just five minutes to impress someone with what AI can do, what would you show them?"

3. “What’s the one thing you do when using AI that you think everyone should copy? Walk us through it.”

4. “What’s an obvious mistake you see people make with AI, and what should they do instead?”

5. “If our viewers should try one AI experiment after this session, what should it be?”

Thoughts?

That’s the mix I’ve got so far, and I feel it covers a good range of topics.

But I’m not you, so what do you think?

If you have suggestions to make this concept even better, please share them:

What Would Make This Better?

The more you share, the more relevant I can make these sessions for all of you.

Thanks for your time!

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:

I was away last week, so we have two weeks’ worth of news to catch up on:

👩‍💻 AI releases

1. Anthropic news:

   1. Claude Code Desktop now has a built-in browser so it can read docs, navigate sites, and interact with web pages directly.

   2. Claude for Teachers gives verified K-12 educators in the US free access to premium features and a built-in teaching skills library.

   3. Cowork is now available on web and mobile, so your background tasks can run on their own without your device needing to be on.

   4. Reflect feature helps you track and visualize your Claude usage so you can better understand your habits.

2. Canva launched Code 2.0, an upgrade to its AI website builder with faster code generation, HTML import, and drag-and-drop editing.

3. Dropbox integrated with ChatGPT and can now help you organize files and conduct multi-step workflows directly inside the app.

4. Google news:

   1. AI Mode in Search now connects to apps like Canva, Instacart, and YouTube Music, so you can finish tasks directly from Search.

   2. Gemini in Chrome rolled out to UK desktop users to help summarize pages, conduct comparisons, and integrate with apps like Calendar, Gmail, and more.

   3. Google Images now features an AI-powered gallery and image generation in Search to mark its 25th anniversary.

   4. Google Photos Video Remix can create shareable video clips with cinematic relighting, background swapping, and other effects.

   5. Google Vids added Gemini Omni editing and personal AI avatars that let you put yourself into video clips.

   6. NotebookLM is now called Gemini Notebook and features native code execution and deeper integration with the Gemini ecosystem.

   7. Waze now includes a motorcycle mode with AI-powered hazard detection plus Gemini conversational search for finding destinations.

5. Meta AI launched Muse Image, an image generator that can edit, restore, and restyle your personal photos using AI.

6. Moonshot AI open-sourced Kimi K3, a multimodal model that gets close to the performance of top-tier frontier models.

7. OpenAI news:

   1. ChatGPT Work connects to Gmail, Google Drive, Salesforce, Slack, and Teams to manage files and run multi-step workflows on your behalf.

   2. GPT-5.6 is a three-tier model family (Sol, Terra, and Luna) with improved intelligence, speed, and agentic capabilities.

   3. GPT-Live voice models let ChatGPT listen and speak at the same time, making voice conversations feel natural with real-time acknowledgments.

8. Perplexity launched Brain, which remembers how you work, switches between AI models faster, and lets you publish websites in one click.

9. PrismML released Bonsai 27B, a powerful AI model that can run directly on phones and laptops for fast, private, offline use.

10. Superhuman launched Auto Drafts 2.0 to write ready-to-send email replies using your inbox, calendar, and personal writing style.

11. Thinking Machines open-sourced Inkling, a multimodal model with controllable thinking effort that can natively process audio, images, and text.

12. xAI news:

    1. Grok 4.5 handles coding, agentic tasks, and knowledge work at close to top-tier model quality while being faster and using fewer tokens.

    2. Grok Automations lets you schedule recurring tasks or create workflows that are triggered by incoming emails.

    3. Grok Build is now open-source, letting developers inspect, customize, and run its coding assistant on their own computers.

Subscribe now

🔬 AI research

1. Apple announced Siri AI, which understands context, takes action across apps, and continues conversations from iPhone to Apple Watch.

2. Roblox previewed Build, a mobile-first tool that lets users create playable games from text prompts without coding.

3. Spotify is testing conversational AI features for Premium users to let them control playback and discover audiobooks, music, and podcasts via text or voice.

📖 AI resources

1. “Claude’s values across models and languages” [RESEARCH]: an Anthropic study of how Claude changes behavior based on language and model version.

2. “CursorBench 3.2” [BENCHMARK]: Cursor’s benchmark that scores coding agents on real, ambiguous tasks from actual development sessions.

🔀 AI random

1. Meta removed its AI image generation from Instagram after three days due to deepfake backlash as it could use anyone’s public photos by default .

2. Suno was hacked to reveal it had scraped over hundreds of thousands of hours of audio from Deezer, Genius, and YouTube.

🤦‍♂️ AI fail of the week

Nature sure works in mysterious ways!

Anthropic’s Fable 5 “will they/won’t they” saga, OpenAI’s new “celestial bodies” family, GLM 5.2, Grok 4.5, and—any day now—Gemini 6.71 Pro A31 Turbo (probably, I don’t even know anymore).

But no matter how good large language models get, there are a bunch of areas where they still struggle…and will continue to do so.

That’s because these issues are baked into how LLMs are built, so a fancier new model may just give you a fancier version of the same problem.

Let’s look at these limitations, why they’re there, and what you can do about them.

1. AI bullshits confidently

Ask AI something specific enough, and it’ll eventually cite nonexistent sources or invent quotes and facts.

What is it?

“Hallucinations” is the term for when AI makes up facts, dates, links, citations, or entire concepts with complete confidence and no hint of uncertainty.

AI would rather tell you that Steven Seagal invented the steam engine during his trip to space than admit it isn’t sure of the answer.

No model is immune to this, although some do better than others:

This has real-world consequences. Here’s a database that tracks almost 1,800 legal decisions affected by AI hallucinations. Here are 12 mainstream cases of hallucinations.

Why does it happen?

Because, during training, language models are rewarded for getting things right and aren’t penalized for wrong guesses.

If you don’t know an answer to a multiple-choice question, you’re still better off picking a random one than leaving it blank, right?

OpenAI’s own research says this:

“…suppose a language model is asked for someone’s birthday but doesn’t know. If it guesses “September 10,” it has a 1-in-365 chance of being right. Saying “I don’t know” guarantees zero points. Over thousands of test questions, the guessing model ends up looking better on scoreboards than a careful model that admits uncertainty.”

Funnily enough, newer models often do worse than their predecessors when it comes to hallucinations.

Grok 4.5 hallucinates at a rate of 54% compared to Grok 4.3’s 25%.

Anthropic found that its newest Mythos 5/Fable 5 model was “…more inclined to attempt an answer than to decline, which improved its correct-rate on knowledge questions but increased the rate at which it fabricated a response when the necessary context is absent.”

So unless we change the entire system of incentives during LLM training, hallucinations are here to stay.

What can you do about it?

Knowledge is power!

You can’t fix hallucinations, but being aware of them should make you more skeptical and suspicious of any specific details coming from a chatbot or AI agent:

• Phrase your requests in a way that lets AI reply with “I’m not sure” instead of making stuff up.

• Verify any critical facts, dates, figures, citations, etc. from LLMs.

• Ask for multiple sources of information (and always check the source yourself).

Subscribe now

2. AI only pretends to reason

AI sure sounds like it knows its stuff, but ask it to make complex calculations or solve a niche riddle, and it may well fail in hilarious ways.

What is it?

We’re all familiar with the now-infamous case of older models failing to count the r’s in “strawberry.”

A more recent example had AI models answer this question:

“I want to wash my car. The car wash is 50 meters away. Should I walk or drive?”

While we all intuitively understand that you can’t wash your car without driving to the car wash, only 5 out of 53 tested models consistently gave the right answer. All others suggested walking at least some of the time, and 33 of them never got it right.

The ARC-AGI range of benchmarks is designed specifically to test this type of fluid intelligence, or “the ability to reason, solve novel problems, and adapt to new situations.”

ARC-AGI puzzles are trivial for humans to figure out, but trip up even the best models.

On the latest ARC-AGI-3 challenge, even the current top-scoring GPT-5.6 Sol (Max) only gets 7.8%.

Why does it happen?

Large language models are built to predict plausible-sounding text, not reliably calculate stuff or track the state of things around them. They don’t have a true world model to lean on that may help them discover solutions to previously unseen problems.

In a 2025 “Comprehension Without Competence” paper, the author writes:

“LLMs often articulate correct principles without reliably applying them—a failure rooted not in knowledge access, but in computational execution.”

LLMs can reliably give correct answers for widely cited riddles and puzzles.

But throw something unique their way, and AI models will struggle to get their bearings.

We end up with a constant game of Whac-A-Mole, where models eventually learn to give the right answers to popular “gotcha” riddles…until another novel challenge comes along and the cycle starts anew.

What can you do about it?

On a broader scale, alternative AI architectures like neuro-symbolic AI may offer the best of both worlds.

For everyday use, you may want to rely on agentic products that equip large language models with harnesses that let them run scripts and other external tools if you care about predictable and reliable answers to math problems, coding challenges, and similar complex tasks.

Subscribe now

3. AI tells you what you want to hear

Every one of your ideas is brilliant. At least AI says so. But is it always true?

What is it?

AI sycophancy is a well-studied phenomenon. AI chatbots tend to roll with your takes. By default, they agree with your arguments, sugarcoat bad news, and otherwise try to please you instead of challenging you.

Why does it happen?

During reinforcement learning from human feedback (RLHF), human raters tend to prefer responses that align with their views, so AI models end up internalizing the fact that raters like agreement more than pushback.

What can you do about it?

While you can’t fully eliminate sycophancy, there are a bunch of things you can try to reduce it. I covered 7 of them here:

Subscribe now

4. AI sounds like, well, AI

In today’s fast-paced world, all AI writing reads the same. It’s not just ChatGPT—it’s every chatbot. In the following section, we’ll delve deeper into its origins, its effects, and its last item on this list of three. No vibes. No hot takes. Just facts.

What is it?

Did that opener instantly set off every “AI slop” alarm bell in your head? I thought so!

We’ve all been conditioned to treat certain recurring tics—like em dashes—as signs of AI writing. Once you start to spot them, it’s hard to unsee them.

Why does it happen?

This one has similar roots to #3.

During finetuning and post-training, human raters prefer texts that sound familiar over anything genuinely novel. This “typicality bias” is passed on to language models, which end up generating repetitive writing patterns in what’s known as “mode collapse.”2

In short, AI picks up the most common turns of phrase and writing conventions and just runs with them.

What can you do about it?

In general, don’t expect AI to ever truly nail the way you write.

Stay in the editor’s seat and don’t outsource your voice to AI. As a writer, you can still use AI for many other things.

But if you’re set on steering chatbots away from “AI slop,” try these:

• Provide curated examples of the kind of writing you’d like to see.

• Give AI a “blacklist” of words, phrases, etc. you explicitly want it to avoid.

• Ask chatbots for multiple outputs with assigned probabilities. This technique, called “verbalized sampling,” gives you a more varied range of outputs.

Subscribe now

5. AI has the memory of a goldfish

The longer your conversation, the more stuff AI starts to forget.

What is it?

Quick: What was Grok 4.5’s hallucination rate? Don’t scroll up!

There’s a non-zero chance you couldn’t remember. After all, it was many words ago.

This happens to AI models, too.

Why does it happen?

Many models now have one-million-token context windows.

But that doesn’t mean they can precisely pick out every detail from that context.

“Context rot” is when the accuracy of AI recall drops sharply as its context fills up.

Even the best models go from near-perfect recall at the start of a chat to a coin flip after 500K tokens.

This is thanks to the transformer architecture behind LLMs, which leads to a “lost in the middle” effect. The first and last messages in a conversation get the most attention. Everything sandwiched in between gets less.

The more tokens in your conversation, the more stuff in the middle of the sandwich that the model might overlook.

What can you do about it?

Companies have already introduced methods like persistent memory, project folders, and .md files to keep key context outside of a single conversation.

From your side, you can do this:

• Start fresh chats for separate tasks or topics, so that each chat has its own relevant context.

• When working with agents like Claude Code and Codex, compact your conversations regularly to keep their context lean.

• Use external folders and tools like Obsidian to maintain a persistent knowledge base for your chatbots and AI agents.

• Lean on whatever existing features help you curate context (e.g. custom instructions, projects, etc.)

Subscribe now

Same architecture, same problems

None of these issues will suddenly be 100% solved with the next model launch.

You can work to reduce them, but you can’t fully train them away.

Not in the current generation of language models, at least.

One day, we may see new AI architectures without these inherent limitations.

Until then, the best we can do is know they exist and learn to work around them.

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

LATEST UPDATE: July 8, 2026

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

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

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

Here are those that did.

Subscribe now

🗣️ 1. Large language models

These are the chatbots and LLMs I turn to.

• 🎁Free:

  • Claude Sonnet 5 [UPDATED - replaced 4.6] for most creative work. Claude has a “fresh” conversational feel and tends to suggest unconventional ideas. Interactive explainers are awesome for visualizing and grasping tricky concepts. The only downside is the strict message limit for free accounts.

  • GPT 5.5 Instant [UPDATED - replaced GPT 5.4] holds up nicely for regular day-to-day chats and low-stakes research.

• 💵Paid:

  • Claude 4.6 Opus is my primary daily driver for everything I do with Claude Code. It runs most of my projects and tasks. (I’m still hesitant to switch to Opus 4.8 due to higher token use and negative public consensus.)

  • GPT 5.5 High [NEW] is my go-to for most heavy research. It tends to do a solid job of surfacing and validating sources, then returning a vetted answer.

🖼️ 2. AI images

Here are my favorite image models.

• 🎁Free:

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

  • GPT Image 2: for leaderboard-topping image generation. It understands context and can render entire pages of text within an image. (I made a swipe file for it with 100 use cases and prompts.)

• 💵Paid:

  • Midjourney: I’ve written dozens of posts about Midjourney over the years. I have honestly not used it much lately, but it remains excellent for artistic exploration and photographic images.

📽️ 3. AI video

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

• 🎁Free:

  • Gemini Omni Flash [NEW - replacing Veo 3.1] is a top-ranked model for context-aware videos with native audio. It’s also great for natural-language video edits. The Flow platform gives you free daily credits to try it.

• 💵 Paid:

  • Seedance 2.0: Still the #1 video model (but Seedance 2.5 is coming) if you want the cutting-edge stuff. (I get to use it on Genspark with my credits.)

🎵 4. AI music

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

• 🎁Free:

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

  • Suno for more creative but natural-sounding results.

Both give you a generous amount of free credits, so take them for a spin to see which one you prefer. These two came out on top in my most recent head-to-head test (sorry, Udio and Sonauto).

Note: For a more skeptical perspective from a professional, see this comment thread.

• 💵 Paid:

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

🔬 5. Research

• 🎁Free:

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

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

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

• 💵 Paid:

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

📈 6. Agents & productivity

• 🎁Free:

  • Codex [NEW]: OpenAI local agent that I now use in tandem with Claude Code. Codex is the most user-friendly way to try local agents without paying (free ChatGPT users get a limited weekly Codex quota).

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

• 💵 Paid:

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

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

Subscribe now

🫵 Over to you…

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

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

Leave a comment

Thanks for reading!

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

• ❤️Like this post if it resonates with you.

• 🔗Share it to help others discover this newsletter.

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

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

Subscribe now

Heads-up: I’m away next week, so the next Sunday Rundown is on July 19.

Happy Sunday, friends!

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. Anthropic news:

   1. Claude Fable 5 has been redeployed after the government lifted its export restrictions, with an improved safety classifier to block harmful requests.

   2. Claude Sonnet 5 is an upgrade to the company’s mid-range model with better coding, reasoning, and tool-use capabilities.

2. Cursor launched a native iOS app in public beta, letting developers build apps and manage coding agents from their phone.

3. Google news:

   1. Gemini Omni Flash and Nano Banana 2 Lite are now available to developers in more places to support conversational video and image creation.

   2. Gemini Spark, the 24/7 AI agent, is now available on macOS to automate desktop tasks, integrating with tools like Canva, Dropbox, and Instacart.

   3. NotebookLM can now turn your sources into 60-second vertical video overviews with narration. (Ultra and Pro subscribers first. Free users later.)

   4. Personal Intelligence can now generate personalized images with Nano Banana based on your data from Gmail, Google Photos, and YouTube.

   5. Take Notes for Me can automatically transcribe Google Meet calls and generate summaries, saving everything to Drive. (Ultra and Pro subscribers.)

4. Runway added Google’s Gemini Omni Flash to its platform, so users can generate and edit videos from images, text, and video inputs.

5. Square now lets restaurants accept orders directly from ChatGPT and Claude, so customers can browse, order, and pay without leaving the chat.

6. X launched an official MCP server that lets AI agents like Claude and Cursor fetch timelines, manage bookmarks, and search posts more easily.

🔬 AI research

1. Google is rumored to be prepping Gemini 3.5 Pro for launch in July, with a two-million-token context window and Deep Think reasoning mode.

📖 AI resources

1. “Getting Started with Loops” [GUIDE]: Anthropic’s look at goal-based agentic loops in Claude. (I wrote a partially related, beginner-friendly take ten days ago.)

2. “Fable 5 vs GPT 5.6 Sol: The Early Results” [VIDEO]: A great dive by AI Explained.

🔀 AI random

1. Netflix will use an AI-generated Gene Wilder voice in its upcoming Wonka’s The Golden Ticket reality show, created with ElevenLabs and the family’s consent.

🤦‍♂️ AI fail of the week

“I’m starting to suspect this octopus isn’t the chess grandmaster we were told it was!”

And then everyone was like, “Yup, Daniel has a point, let’s scrap the idea.”

And all AI labs cancelled their web search implementations.

And nobody asked a chatbot to research anything ever again.

Remember that? Good times!

Anyway, let’s return from La-La Land back to reality, where over 50% of all search engine volume is now driven by AI assistants.

I’m not immune to this, by the way.

I now turn to chatbots and agents for my research far more often than to Google.

But did you know that the language of your AI chats might inadvertently limit your perspective?

Let me show you what I mean and how you can do better.

The “language lens” problem

You might think that when you ask a chatbot a question, it diligently combs the entire Internet, gathering information from all around the world to give you a global synthesis.

But nope.

That’s not what happens.

Instead, the chatbot searches the Internet as you would, in your language.

This seems like a minor nitpick, but this approach often overlooks local perspectives.

Let’s try a simple search for Denmark news in ChatGPT:

Prompt: Summarize the five biggest news stories about Denmark of the past week. Just a concise one-sentence bullet for each item. Make it easy to skim.

Here’s what that gives us:

Seems solid.

But now watch what happens when I add a simple language anchor:

Prompt: Summarize the five biggest news stories about Denmark of the past week. Just a concise one-sentence bullet for each item. Make it easy to skim. Search in Danish. Reply in English.

Here’s the new list:

While there’s definitely some overlap (record heat, airport drones), the Danish-anchored perspective is different in several ways:

1. The overall mix is more local and inside-out (migration laws and health reform) instead of geopolitical (Greenland and Ukraine military aid).

2. The mix of sources is exclusively Danish, compared to the original’s Reuters, AP News, and The Copenhagen Post (English-language newspaper for internationals).

3. The search took almost one-and-a-half minutes (compared to just 26 seconds in English), indicating that ChatGPT had to dig harder to surface local sources.

What this shows is that the chatbot’s answers are subtly tainted by your query language. (Unless you proactively counteract it.)

That’s the language blind spot in action.

Uh...so what?

This might feel like a non-issue if all you’re after is a quick at-a-glance view.

But there are situations where a local perspective is genuinely helpful.

The hidden superpower of large language models isn’t that they can summarize stuff. It's that they can search and understand parts of the web you'd never explore yourself.

Language barriers largely don’t exist for chatbots.1

Think of what that means.

You can learn what Chinese media outlets actually write about a trade dispute, not a summary of China’s perspective from Reuters.

You can find authentic reviews of a Japanese product directly from Japanese sites.

Or consult a German health forum about a treatment before going to a local clinic.

Or discover hidden gems in a Brazilian town that locals actually prefer to visit.

And so on.

Subscribe now

The fix: ask AI to search locally

The trick is to explicitly ask the chatbot to search local sources in a specific language, as you already saw with my Danish news example.

Use this:

Prompt: [My question or request]. Search [location] sources in [language]. Reply in [my language].

Add this location/language line to any research prompt, and your chatbot will pull info from local sources and translate the findings for you.

This approach works with virtually any query where a local perspective matters.

Limitations & caveats

Of course, even this kind of local-aware search isn’t perfect.

Here are some limitations:

• Online only: You’re still at the mercy of what’s actually available digitally. Local TV channels, magazines, and newspapers are out of your reach.

• Accessibility: Not every site is crawlable by AI agents and chatbots, so even if the content exists, they might not find it.

• Popularity bias: Chatbots are likely to surface well-ranked results rather than lesser-known or smaller sites.

• Source quality: Chatbots may include results from questionable sources or state-controlled outlets without making this transparent.

• Verification friction: With results in your language, you can go to the source and double-check AI’s answers. With foreign-language sources, this gets trickier.2

Still, inviting a local perspective is almost always better than settling for vanilla results.

Subscribe now

Bonus: My “Global Pulse” skill (+prompt)

That above one-line addition will already upgrade your research.

But it gives you a single search in a single language.

I’m a big fan of what Ground News does for news coverage, showing how different sources cover the same story.

I wanted something like that, but across languages and not only for news.

So I built a Global Pulse skill for Claude Code (or any other local agent).3

Hand it your research query and a list of countries to cover.

Global Pulse then searches local sources, compares perspectives side by side, uncovers unique local-only insights, and much more.

You get a polished HTML report that shows you:

• A dedicated per-country summary with key findings and local scoops

• A clickable list of sources per country for further exploration

• A cross-country synthesis with the full range of perspectives

Here’s a one-country sample teaser:

For those who don’t use AI agents or want a simpler solution, there’s also a copy-paste prompt (750+ words) that works in any chatbot with web search.

Grab the “Global Pulse” skill-and-prompt combo here:

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 releases

1. Anthropic brought the Claude Desktop app to AWS, Google Cloud, and Microsoft Foundry. (Read my recent primer on Claude Code Desktop features.)

2. Figma launched a suite of new AI features like generative plugins, Weave tools, shader fills, and agent upgrades to help designers get more out of their canvas.

3. Google news:

   1. Gemini 3.5 Flash now has built-in computer use, letting AI agents see, click, and type across browser, desktop, and mobile environments.

   2. Gemini in Chrome now lets you select any text or image on screen and send it to Gemini so it can work on specific elements instead of the entire page.

   3. Google Finance is out of beta with AI-generated market briefings and portfolio tracking. (It also comes with a new Android app.)

4. Mistral released OCR 4, a document extraction model that can read PDFs, presentations, and Word docs in 170 languages.

5. Perplexity launched Computer for Counsel, a legal AI agent that connects with external tools to help with contract drafting, document review, and research.

6. Runway launched Agent 2.0 that can turn a single prompt into marketing briefs and campaign assets for multiple platforms.

7. Sakana AI launched Fugu, a multi-agent system that automatically picks the best AI models for each part of your task to deliver top-tier results.

🔬 AI research

1. Anthropic launched Claude Tag that lets teams tag Claude in Slack to work on tasks, answer questions, and surface relevant updates. (Beta for Enterprise users.)

2. ByteDance previewed Seedance 2.5, a video model that can generate native 30-second 4K clips and handle up to 50 reference inputs.

3. OpenAI previewed the new flagship GPT-5.6 Sol model and its smaller siblings, Terra (everyday work) and Luna (fast and cheap). (Limited release for now.)

4. xAI began testing Grok 4.5 in private beta at SpaceX and Tesla, with early results allegedly competitive with frontier models like Claude Opus.

📖 AI resources

1. “AI Watchdog” [TOOL]: searchable database by The Atlantic that lets anyone look up artists, musicians, channels, etc. that were used to train AI models.

2. “FFASR Leaderboard” [BENCHMARK]: new benchmark to test speech-to-text agents on real-world challenges like accents, noise, and multiple languages.

🔀 AI random

1. Getty Images signed a multi-year deal with OpenAI to show licensed photos and videos in ChatGPT search results.

🤦‍♂️ AI fail of the week

Nailed it, GLM-5V-Turbo! That’s GPS-level precision right there.

📹 Live & Learn #7: AI vision capabilities

This week’s Live & Learn was all about testing how well AI models “see” stuff:

Live & Learn is taking a hiatus for the next few weeks (vacation, family visits, etc.), but it’ll be back with a vengeance in August.

Stay tuned!

This week, I flipped the script: Can AI models “see” images?

I put five free-tier chatbots through four vision challenges

If you’ve ever wondered whether your chatbot of choice can actually understand what you show it, this one should give you a clear answer.

(Spoiler alert: only one model found Waldo, and it’s one I didn’t expect.)

Tools tested & scorecard

Read more