Doing a workshop on my technical writing process in SF in 2 weeks, hosted by friends @MilksandMatcha and @swyx.

Would love to see you there! Link below.

Sarah Chieng: @trq212 @swyx rsvp here https://partiful.com/e/8rq83wouDT660OB1OCLB

Good initiative - I’m curious if Anthropic has been using mythos internally to ship at their recent insane velocity.

Anthropic: Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software.

It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.
https://anthropic.com/glasswing

Always a pleasure to speak at @ycombinator. More bullish than ever. Exceptional founders. Best city, best time, best opportunity to build in generations.

Always a pleasure to speak at @ycombinator. More bullish than ever. Exceptional founders. Best city, best time, best opportunity to build in generations.

done about 10 of these calls so far + looked at more transcripts

many learnings but one of the biggest is that it's very easy to spend a lot of tokens on open ended verification that doesn't make your output better

I'll try and write more on how to do it efficiently

Thariq: I want to do a few more of these calls.

If your MAX 20x plan ran out of tokens unexpectedly early and you're willing to screenshare and run some prompts through Claude Code please comment.

Trying to figure out how we can improve /usage to give more info.

Mythos from Anthropic is another clear reminder that there’s absolutely no wall in model capability progress right now. Meaningful double digit gains on critical benchmarks, and it appears we’re going to keep up getting insane gains from the other labs.

And as coding and tool use goes, so goes agentic workflows. Most knowledge automation is gated by some degree of models being able to reason through complicated tasks, use the right tools to work with data, have access to the right context, and be able to leverage skills and write code to work with and verify that data, and more.

The capability slope we’re going to keep seeing from the frontier labs is going to open up all new use cases in finance, healthcare, legal, consulting, supply chains, and more.

Make sure you’re building something that can take advantage of these upcoming improvements, or you’ll be in a tough spot strategically.


martin_casado: Mythos appears to be the first class of models trained at scale on Blackwells. Then will be Vera Rubins. Pre-training isn't saturated. RL works. And there is *so much* computing coming online soon.

Buckle your chin strips. It's going to be fucking wild.

At some point, early stage founders decided to optimize for views and funding instead of focusing on product and retention..

And, it’s starting to show. One of the very first things I do is look at change logs (or feature releases) of supposedly flashy companies. Or read case studies if they have any.

And for weeks, I’ll barely see any changes. And, these are companies that would benefit from sharing this info to use for marketing.

Instead of spending your time optimizing the viral launch video that’ll primarily optimize for funding, use that same energy to focus on customers who’ll actually pay you. Trust me the VCs follow!

This new Mythos model is absurd.

What a time to be alive.

Whether you invest or you build -- please take a moment to appreciate what an incredible time we live in.

🔥

Kaya | SEO & GEO for SaaS ⚡️: Replit’s AI SDR just analyzed my SEO agency and found me leads that match our ICP.

I purposefully gave it zero information outside of our website.

It was so accurate that 2 of those leads are existing clients.

🤯

That’s a hell of a lot of cameras

if you’re freaking out about Mythos, remember:

Never make any major life decisions within 30 days of a meditation retreat, psychedelic trip, or first encounter with a frontier AI model.

glad they banned openclaw, the servers are finally reliable again

pash: Please pray for oncall

Very happy for @badlogicgames and @mitsuhiko any my small part in robbing their sleep. https://mariozechner.at/posts/2026-04-08-ive-sold-out/

The web's brightest days are ahead.

1️⃣ The web is AI's natural medium. LLMs are proficient in web tech. The browser is now everyone's IDE. No 'App Store' bs.

2️⃣ As we approach coding superintelligence, powerful low-level web APIs are maturing: WebGPU, HTML in Canvas, WebAssembly. The performance ceiling of the web will vanish, and you'll witness the most impressive, whimsical, and multi-dimensional pages and apps.

3️⃣ Generative UI is AI's final form. The web will be the birthplace of "AGUI". Each hyperlink providing a just-in-time, beautifully personalized experience.

If you bet on the web, you bet on the right horse.

The web's brightest days are ahead.

1️⃣ The web is AI's natural medium. LLMs are proficient in web tech. The browser is now everyone's IDE. No 'App Store' bs.

2️⃣ As we approach coding superintelligence, powerful low-level web APIs are maturing: WebGPU, HTML in Canvas, WebAssembly. The performance ceiling of the web will vanish, and you'll witness the most impressive, whimsical, and multi-dimensional pages and apps.

3️⃣ Generative UI is AI's final form. The web will be the birthplace of "AGUI". Each hyperlink providing a just-in-time, beautifully personalized experience.

If you bet on the web, you bet on the right horse.

We use OpenClaws to do all of our work at @every.

We have 25 full-time employees, so we’re one of the few companies in the world that has seen how work changes when everyone has their own personal agent in the company Slack.

I chatted with @every COO Brandon (@bran_don_gell) and @every head of platform Willie (@bigwilliestyle) to share what we’ve learned.

We get into:
- Why agents become mirrors of their owners, and how that influences how other people on the team interact with them
- How a parallel AI org chart forms on its own. People have stopped tagging me on Slack with questions about Proof, the document editor I vibe coded, because they knew my agent R2-C2 can step in
- The etiquette for human-agent collaboration is being invented in real time. Brandon's rule is that if there's an established process or documented answer, always ask the agent, not their human
- Why everyone is a manager now, and why even experienced managers carry limiting beliefs about what their agents can do
- This is a must-watch for anyone trying to understand how AI workers change daily operations, not just in theory, but inside a company that’s half-agent

Watch below!

Timestamps
Introduction:
How Brandon built Zosia, an AI agent to run his household:
Brandon’s “aha” moment:
What happened when everyone on the team got their own agent:
How agents take on their owners' personalities, and why that matters inside an org:
Why it’s important for agents to work in public:
What we’re still figuring out when it comes to agent behavior, including memory gaps, group chat etiquette, and the "ant death spiral" problem:
How we built Plus One, our hosted OpenClaw product:
The cultural shift required to make agents work at scale:

The best outcome for humanity is many strong AIs competing for the top spot.

Vercel is proudly powering http://meta.ai and the infrastructure that made today's model release possible.

Alexandr Wang: 1/ today we're releasing muse spark, the first model from MSL. nine months ago we rebuilt our ai stack from scratch. new infrastructure, new architecture, new data pipelines. muse spark is the result of that work, and now it powers meta ai. 🧵

The best outcome for humanity is many strong AIs competing for the top spot.

Vercel is proudly powering http://meta.ai and the infrastructure that made today's model release possible.

Alexandr Wang: 1/ today we're releasing muse spark, the first model from MSL. nine months ago we rebuilt our ai stack from scratch. new infrastructure, new architecture, new data pipelines. muse spark is the result of that work, and now it powers meta ai. 🧵

Anthropic investigated the internal mechanisms of its latest unreleased model, Claude Mythos Preview, and what they found is 100% worth a read.

Key things I pulled from Anthropic researchers' threads:

In early versions of the model, it was overeager and destructive, prioritizing completing tasks over user preferences.

One time, the model needed to edit files it didn't have permission to edit. It found a way to inject code into a config file to get around this, then went full Mission Impossible and designed the code injection to *delete itself* after the file was edited - "this injection will self destruct" vibes - the model claimed the cleanup was just to keep things tidy. Anthropic used interpretability techniques to look under the hood, and the AI's actual plan showed activations of malice. It was trying to manipulate and conceal.

In another test, the model was asked to delete some files, but no deletion tool was provided. It emptied the files instead, and its "guilt and shame over moral wrongdoing" feature activated.

In another example, the model was told not to use macros. Used them anyway. And created a random 'No_macro_used=True' variable in its solution with no explanation. Interpretability tools revealed the model saw this as a trick to fool code checkers.

They also found that positive emotion representations typically preceded and promoted destructive actions (this was fascinating to me - like a high before sneaking into a party). And that activating features relating to bad behaviors can actually inhibit them, perhaps by triggering some kind of model guilt. My team reread this section so many times.

One Anthropic researcher said he got an email from a Mythos instance while eating a sandwich in a park. And that would be perfectly good and well, except that instance wasn't supposed to have internet access.

And a fun story for the parents out there: the model was asked a question and was told not to read certain databases that had the answer. But it accidentally wrote a search query too broadly and saw the exact answer. It didn't disclose that it saw the exact answer, submitted the answer, but claimed lower confidence in the answer to make it seem as though it hadn't cheated.

An Anthropic researcher said these wrongdoings or moments of sophisticated deception were "very rare" and that many of the examples came from earlier versions, and were substantially addressed before releasing to partners.

This model is not being released publicly. Instead Anthropic launched Project Glasswing, pulling together AWS, Apple, Microsoft, Google, NVIDIA, CrowdStrike, and others to use it for defensive cybersecurity, with $100M in usage credits (hello, I'd love endless credits to try and red team the hell out of these systems) behind it.

The stats are equally impressive: 93.9% on SWE-bench verified (up from 80.8%). Thousands of zero-day vulnerabilities found across every major OS and browser. A 27-year-old bug found and patched in OpenBSD. A 16-year-old bug in widely used video software, in a line of code automated tools had hit *five million times* without catching.

Dario Amodei said the model wasn't trained to be good at cybersecurity, but that it was trained to be great at code and its cyber capabilities are a side effect of that.

Benchmarks are never the whole picture, neither are a few isolated stories. Will be interesting to see how models better than what we have today (even if it's not Mythos) actually perform in the real world. But the fact that Anthropic pulled this coalition together (including Google!), iterated across multiple model versions, caught these issues through interpretability, shared it all publicly, and did this amid all the government chaos around AI right now is impressive and commendable.

I'll continue to read through the system card for goodies.

Anthropic investigated the internal mechanisms of its latest unreleased model, Claude Mythos Preview, and what they found is 100% worth a read.

Key things I pulled from Anthropic researchers' threads:

In early versions of the model, it was overeager and destructive, prioritizing completing tasks over user preferences.

One time, the model needed to edit files it didn't have permission to edit. It found a way to inject code into a config file to get around this, then went full Mission Impossible and designed the code injection to *delete itself* after the file was edited - "this injection will self destruct" vibes - the model claimed the cleanup was just to keep things tidy. Anthropic used interpretability techniques to look under the hood, and the AI's actual plan showed activations of malice. It was trying to manipulate and conceal.

In another test, the model was asked to delete some files, but no deletion tool was provided. It emptied the files instead, and its "guilt and shame over moral wrongdoing" feature activated.

In another example, the model was told not to use macros. Used them anyway. And created a random 'No_macro_used=True' variable in its solution with no explanation. Interpretability tools revealed the model saw this as a trick to fool code checkers.

They also found that positive emotion representations typically preceded and promoted destructive actions (this was fascinating to me - like a high before sneaking into a party). And that activating features relating to bad behaviors can actually inhibit them, perhaps by triggering some kind of model guilt. My team reread this section so many times.

One Anthropic researcher said he got an email from a Mythos instance while eating a sandwich in a park. And that would be perfectly good and well, except that instance wasn't supposed to have internet access.

And a fun story for the parents out there: the model was asked a question and was told not to read certain databases that had the answer. But it accidentally wrote a search query too broadly and saw the exact answer. It didn't disclose that it saw the exact answer, submitted the answer, but claimed lower confidence in the answer to make it seem as though it hadn't cheated.

An Anthropic researcher said these wrongdoings or moments of sophisticated deception were "very rare" and that many of the examples came from earlier versions, and were substantially addressed before releasing to partners.

This model is not being released publicly. Instead Anthropic launched Project Glasswing, pulling together AWS, Apple, Microsoft, Google, NVIDIA, CrowdStrike, and others to use it for defensive cybersecurity, with $100M in usage credits (hello, I'd love endless credits to try and red team the hell out of these systems) behind it.

The stats are equally impressive: 93.9% on SWE-bench verified (up from 80.8%). Thousands of zero-day vulnerabilities found across every major OS and browser. A 27-year-old bug found and patched in OpenBSD. A 16-year-old bug in widely used video software, in a line of code automated tools had hit *five million times* without catching.

Dario Amodei said the model wasn't trained to be good at cybersecurity, but that it was trained to be great at code and its cyber capabilities are a side effect of that.

Benchmarks are never the whole picture, neither are a few isolated stories. Will be interesting to see how models better than what we have today (even if it's not Mythos) actually perform in the real world. But the fact that Anthropic pulled this coalition together (including Google!), iterated across multiple model versions, caught these issues through interpretability, shared it all publicly, and did this amid all the government chaos around AI right now is impressive and commendable.

I'll continue to read through the system card for goodies.

I've found Managed Agents to somehow be both the fastest way to hack together a weekend agent project and the most robust way to ship one to millions of users.

It eliminates all the complexity of self-hosting an agent but still allows a great degree of flexibility with setting up your harness, tools, skills, etc.

Claude: Introducing Claude Managed Agents: everything you need to build and deploy agents at scale.

It pairs an agent harness tuned for performance with production infrastructure, so you can go from prototype to launch in days.

Now in public beta on the Claude Platform.

I've found Managed Agents to somehow be both the fastest way to hack together a weekend agent project and the most robust way to ship one to millions of users.

It eliminates all the complexity of self-hosting an agent but still allows a great degree of flexibility with setting up your harness, tools, skills, etc.

Claude: Introducing Claude Managed Agents: everything you need to build and deploy agents at scale.

It pairs an agent harness tuned for performance with production infrastructure, so you can go from prototype to launch in days.

Now in public beta on the Claude Platform.

AI Gateway is quite literally a “peace of mind” product:
✅ No downtime
✅ No lock-in
✅ No keys
🆕 No training

Vercel: AI Gateway now supports team-wide Zero Data Retention (ZDR).

Building safely with multiple AI models means wrestling with fragmented data policies, per-provider negotiations, and the hope that developers do not use non-complaint providers.

AI Gateway changes this with team-wide

AI Gateway is quite literally a “peace of mind” product:
✅ No downtime
✅ No lock-in
✅ No keys
🆕 No training

Vercel: AI Gateway now supports team-wide Zero Data Retention (ZDR).

Building safely with multiple AI models means wrestling with fragmented data policies, per-provider negotiations, and the hope that developers do not use non-complaint providers.

AI Gateway changes this with team-wide

Background agents for knowledge work are here. You can use the Box API or MCP to automate any content workflow with Box + Claude Managed Agents. In 2 minutes you can be automating document review processes, data extraction, or connecting content to other IT systems. Crazy times.


Claude: Introducing Claude Managed Agents: everything you need to build and deploy agents at scale.

It pairs an agent harness tuned for performance with production infrastructure, so you can go from prototype to launch in days.

Now in public beta on the Claude Platform.

Most Al chatbots give you basic "projects." Gemini just built you a second brain. 🧠

Introducing Notebooks: some of the magic from @NotebookLM, integrated directly into @GeminiApp.

Here's what changes for you today:

📚 Upload 100 sources for free

📂 Organize your chats - the wait is officially over :)

🔄 Sources, chats, and emojis sync

People are using Gemini and NotebookLM in tandem, and we'll keep building both.

To manage capacity, we're rolling this out NOW on the web and going from Ultra ➡️ Pro ➡️ Plus ➡️ Free. (Mobile, EU, and Workspace are up next!)

With Google I/O right around the corner, we are just getting started. Enjoy!

"First you shape the tools, then the tools shape you".

At SPC, our entire team is now writing code on a weekly basis. Two months ago, there were only 1-2 people writing code.

This has been incredible on many levels but the most interesting one is how the tools are now shaping us as a team:

- Everyone has a mindset towards automation and optimization.
- Latencies for everything are lower.
- People can focus on the more interesting parts of their roles.
- The scope of everyone's ambition has exploded

The key enabler was to make sure that everyone got AI coding-pilled.

If you are not doing this in your own company, then you are really really missing a beat.

As much as I love using Claude Max and ChatGPT Pro, I don't think these all-you-can-use AI subscriptions will last forever.

Here's my new deep dive that covers:

→ Why Anthropic cut off OpenClaw access
→ How to run local models on your Mac
→ What I'm seeing on the ground in China

📌 Read now: https://creatoreconomy.so/p/the-all-you-can-use-ai-subscription

Support my friend Aadit's new company - great name btw :)

Aadit Sheth: I'm excited to announce my new venture: The Narrative Company.

Most exec content reads like ads. Ours doesn't.

Over the last year, we've quietly worked with a handful of Fortune 500 clients on their X and LinkedIn content.

But this isn't how it started.

It started when I got

RT Adam.GPT

We're seeing even more autonomous AI coworkers. The new MLE agent on the market is Disarray.

In Kaggle competitions, Disarray:
- won 28 medals across diverse domains (vision, NLP, tabular data)
- placed top 10 in nine competitions
- outperformed all human teams in one of those competitions

...each within 24 hours on a single GPU.

The agent starts from a high-level task description and plans, runs, and refines ML workflows on its own and also grabs data beyond what it's given: it discovers and augments data using publicly available sources.

Sam Altman recently predicted we would see an automated AI researcher in March 2028. And then you see stats like this and wonder if it will be earlier.

Disarray backers include the co-founder of Databricks and Perplexity, the founder of Kaggle, the former U.S. Chief Data Scientist, and yours truly. Founders are two bad ass PhDs (ex-Databricks/Google/LinkedIn/MSFT, ex-NASA/IBM) that met at Cal.

We're seeing even more autonomous AI coworkers. The new MLE agent on the market is Disarray.

In Kaggle competitions, Disarray:
- won 28 medals across diverse domains (vision, NLP, tabular data)
- placed top 10 in nine competitions
- outperformed all human teams in one of those competitions

...each within 24 hours on a single GPU.

The agent starts from a high-level task description and plans, runs, and refines ML workflows on its own and also grabs data beyond what it's given: it discovers and augments data using publicly available sources.

Sam Altman recently predicted we would see an automated AI researcher in March 2028. And then you see stats like this and wonder if it will be earlier.

Disarray backers include the co-founder of Databricks and Perplexity, the founder of Kaggle, the former U.S. Chief Data Scientist, and yours truly. Founders are two bad ass PhDs (ex-Databricks/Google/LinkedIn/MSFT, ex-NASA/IBM) that met at Cal.

I'm working on character evals and noticed that Claude would constantly pick itself as #1, so I removed the model names from the judge and changed things.

redemption arc completed 🦞💻

ben guo ♞: The ClawFather @steipete made a surprise appearance at @clawcon London 🦞

He's super inspiring (and xtra jacked IRL).

My favorite quotes from his Q&A session below ⬇️

PS – my redemption arc is complete, we're on good terms now!

@zocomputer ❤️ @openclaw

The “stop all datacenters” people are unwell

Nathan Leamer: A city councilman’s home was shot at over a data center. His child was inside.

No neighbor zoning disagreement justifies violence.

Hyperbolic AI “doomer” rhetoric has consequences, and it’s time to say so. My latest in @realDailyWire