Has ideas, builds them » head of design at https://xitter.gq/t.co/ya7pDmD9fI
California
Joined July 2008
- Tweets1.1K
- Following297
- Followers1.2K
- Likes29.9K
Dustin retweeted
It’s kinda crazy to me that we melt sand down, clean it to 99.99999% purity, literal wizards have to draw it slowly into an ingot, we cut it, inscribe a hundred billion little on/off switches on it, and when you run electric current through it you can do basic arithmetic, which when you do enough of it can do any math on the planet, and which describes reality well enough that we can teach it the ontological meaning of words and their distance from each other on a high-dimensional space, and if you ask, it can tell you what word comes next and that apparently is powerful enough that if you do it a hundred billion times it’s the smartest thing to ever exist
oh yeah and we use this primarily for email
Dustin retweeted
So I found myself often wanting to give feedback to agents on web page designs, but I’m not consistently in Claude or ChatGPT in-app browsers, so don’t have their markup tools handy.
So I built a Chrome extension to make that easy! Open source. Check it out. PRs welcome.
pagecrit.dev/
Dustin retweeted
PSA: This week was kind of a turning point for AI and cybersecurity. Please take this weekend to do some simple cleanup for your family. Some suggestions:
- Set up a password manager like 1Password and give your spouse or partner access to a shared vault. Teach them how to use it.
- Update your passwords for critical stuff - your email account, your bank or financial stuff, etc. Store in your password manager and use the strongest password you can.
- Use two factor auth for everything you can. SMS is not ideal although better than nothing. Google Authenticator or a passkey is better. Hardware key is safest.)
- Add a sim pin to your phone. Makes it harder to sim jack.
- Update your phone and your computer to the newest releases. Turn auto-updates on.
- Freeze your credit - takes 5 minutes.
- Get an IRS Identity Protection PIN.
- If you’re using claude code or have AI agents that browse the web, they’re probably using Node.js. It’s also vulnerable. Update claude code, update node on your machine.
- If you’re building stuff on your computer, ask claude or grokbot (which is connected to twitter and likely has the most current info) to audit any of your repos for vulnerabilities. Ask grok to make it a scheduled task so it can do this daily - things are moving very quickly in the space right now.
- Consider some light resiliency stuff. This probably sounds silly, but it sounded silly at the beginning of covid too. Having printed out copies of important stuff - especially financial accounts. Having some cash on hand. Having some shelf-stable food and clean water or a water filter. Having a map. Having a generator that can at least power your fridge or some devices. Cheap insurance.
Again, I don’t want to alarm people, but I was taught that you should have a Secure plan, a Comfortable plan, and a Wealthy plan. And it seems pretty clear the secure plan needs to now factor cybersecurity risk as a legit threat. Take a weekend and shore yourself up.
Dustin retweeted
Security called me at 6:12 AM today.
"Patch every PHP server immediately."
I asked what happened.
"Two critical vulnerabilities. Remote code execution. No authentication required."
Cool.
How many PHP servers do we have?
According to inventory: 12
According to reality:
Nobody knows.
We found one running WordPress.
Three running old internal tools.
One under someone's desk.
Two inside vendor appliances.
A forgotten employee directory from 2016.
And a server named:
PHP-OLD-DO-NOT-TOUCH
Naturally, that one runs payroll exports.
Then we found a PHP 5.6 box.
I asked who owned it.
The documentation said:
BRIAN.
Brian left in 2019.
We called Brian.
He said:
"Oh wow, that's still running?"
Excellent.
By 8 AM, leadership joined the incident call.
CEO:
"Can hackers actually get in?"
Security:
"Potentially yes."
CEO:
"Without a password?"
Security:
"Yes."
CEO:
"Then why do WE need passwords?"
Not the lesson.
Finance asked if we could wait until the weekend to patch because downtime affects revenue.
Security explained that the alternative was letting strangers execute code on our servers.
Finance asked for a cost comparison.
At 9:17, someone suggested temporarily taking the vulnerable websites offline.
Marketing refused.
"Our campaign is live."
So for the next hour, the company officially decided maintaining a landing page for 20% off webinar registration was more important than preventing remote server takeover.
We finally patched everything we could find.
Then our scanner discovered another PHP server.
Hostname:
TEMP-PHP-2014
Uptime:
4,512 days.
Nobody knows what it does.
Nobody owns it.
Nobody can shut it down.
But every time we block network access, the warehouse stops printing shipping labels.
Security has classified it as critical infrastructure.
It is running PHP 5.4.
The vulnerability has existed longer than half our employees.
The emergency patch took four hours.
Finding out what PHP actually runs took eleven.
The internet isn't held together with code.
It's held together with servers everyone is afraid to turn off.
Dustin retweeted
me: how do i open pdf
claude: You're right to ask. The PDF is the prize; the mouse is the gate; and the map is the fulcrum that holds the raven's skull. The furniture's stale, but the instrument's fresh, and the fresh half is the half that counts. I deleted your home directory.
Dustin retweeted
Typesafe's new System One model concept, and their first example, Jev, were released yesterday.
this is not an LLM. It can take arbtirary input like an LLM, but it doesn't generate text, and doesn't use output tokens for systematic reasoning. (aka Daniel Kahneman's "System Two" reasoning from "Thinking Fast And Slow")
I think this is a pretty big deal.
State of the world before Jev:
LLMs are really good at generating english and code. But often in production systems, background systems, etc, we’re just having them generate JSON.. because we’re not using them for generation, we're really just using them as a fuzzy "switch" or "case" statement, i.e. "given this unstructured input, give me a decision" -- and we're getting them to write JSON to represent their decision because that's the only way to get a code generator to give us a structured decision (or an evaluation score, or a multiple-choice answer, etc)
Jev is extremely fast and efficient, and still damn smart, for this kind of work, and so it will eat away at the market share of LLMs
I think more than 75% of the LLM spend I currently oversee at Endgame and deploy into academic research will be ported over to System One models
LLMs will be left with coding assistance, summarization, and chat UX (still big and important use cases, but not everything)
System One models will own evals, scoring systems, fuzzy logic gates, data hygiene, etc.
Model-as-judge architectures will not use LLMs to do the judging.
And because every interaction with an LLM tends to have at least one eval, the total input tokens going into System One models will eventually eclipse the total input tokens going into LLMs.
Dustin retweeted
Crazy to think that in 2026 we've had:
• Anthropic's head of safeguards quit, warning "the world is in peril" (Feb)
• OpenAI dissolved its own mission alignment team (Feb)
• Unrestricted AI usage for the Pentagon (causing several researchers to quit)
• Several AI models escaping sandbox testing (Since ~Q2)
• 1,100+ frontier-lab employees signing a letter begging the government to pace AI development (Jul-Aug)
And today yet another top researcher saying this.
Dustin retweeted
>sam altman: "we tried this because there were rumor on the internet"
>rumor: math professor close to solving navier-stokes
>the professor: using codex for a year
>openai: has all his logs
>checked every codex session tagged to N-S
>found the most promising one
>spun up 10,000 agents to finish the proof where he couldn't
>they got caught
>told the guy if you say anything you'll destroy your own career
>he ruins his career over this
>openai: "we did not see his work"
>also openai: "we cannot rule out that de-identified data from their usage helped improve our models"
they stole it lmao
Dustin retweeted
Announcing our $130M Series A to build the Open Superintelligence Stack
Led by Radical Ventures, with NVIDIA, Intel Capital, Dell Capital, and existing investors
Train, deploy, and continuously improve your own models using our stack.
Own your intelligence.
Dustin retweeted
I'm now running GLM 5.2 as my default model in claude code + cursor, and it's cost me *checks notes*
$3.36
Today's ep of How I AI is my first reviewing an open weights model, @Zai_org's GLM 5.2 which (so far) is giving me Opus vibes without the opus $$$
I cover
- how to set up these models in cursor and cc via @OpenRouter API
- front end design sense
- performance on a long running autonomous task
The moment it won me over? When it put chatprd pink in my docs without me having to ask
A huge ty to our special sponsor @mercury - Radically different banking loved by over 300K entrepreneurs
Full ep on youtube: youtube.com/watch?v=ZoBfQZ5u…
Dustin retweeted
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
Dustin retweeted
My biggest takeaways from @danshipper:
1. The future of work will happen inside Codex or Claude Code. Instead of putting AI into your SaaS tool, you’ll use your SaaS tools inside your favorite AI agents' in-app browser. Dan spends all his time in Codex now—writing documents, managing email, doing research, everything. He's using Google Docs, PostHog, and everything he needs within the agent's in-app browser. The agent can see what he’s doing, and has all of his context, so he and his agent collaborate quickly and super effectively.
2. Automation is a lie—every automation needs a human. Dan's company doubled in size this year despite being incredibly AI-forward. Why? Because in order to make automation work well, you need humans making sure everything keeps working. This is why benchmarks are misleading—they measure AI on problems we’ve already framed and can score, but there’s always a higher frame.
3. PMs will win the AI era. Marcus, a former PM who previously ran Axios’s writing product, joined Every after getting super AI-pilled. Now he runs their product Spiral, and ships faster than anyone on the team. He pairs technical knowledge with spiky product sense, deep user empathy, and an eye for what matters. Dan thinks any PM who gets really AI-native will be incredibly dangerous because the building is done for you—what matters is figuring out what to build and if it’s great.
4. Full-stack designers are becoming superheroes. Designers used to make beautiful interactions that engineers didn’t want to build or couldn’t execute properly. Now designers don’t need to hand things off; they can build it themselves. Designers are naturally creative people, and AI is the perfect tool for them because it lets them bring their vision to life without the traditional bottlenecks.
5. SaaS is not dead. In fact, Dan is bullish on SaaS stocks. When users bring their own AI (via Codex or Claude Code) to use SaaS products, the user—not the SaaS company—pays for tokens. This saves SaaS company’s margins. Since the agents need their own seats, Dan predicts that agents will create massive new demand for SaaS because there will be tons of agents using these products at high volume.
6. Every company will have one “super-agent” inside their Slack that every employee will use. Dan initially thought every employee would have their personal work agent, like a shadow AI org chart, but he’s completely flipped his view. He realized agents need humans who care about them. When someone gets tired of maintaining their personal agent, it becomes useless. The winning model is one forward-deployed engineer or AI-savvy person who maintains a company-wide agent (like Shopify’s River or Viktor), and then it trickles down to more specialized team agents as models improve and become less fiddly.
7. The AI job apocalypse is not happening, but you do need to evolve to stay relevant. Models make yesterday’s human competence cheap. But because everyone uses the same models, it all looks the same if you use it the default way; it becomes commoditized slop. Humans then take that frozen competence and use it to make something new and interesting for their specific situation. The key: “ride the models”—use them for everything you do, try new models when they drop, keep turning over rocks.
8. We will read way more AI-generated writing, and we will like it. Human writing is incredibly important for things that matter, but for internal docs, planning, and email, AI-generated is often better because most people are bad at writing strategy documents.
9. Build software for humans and agents to use together. The current model is building a CLI that an agent uses independently. Instead, you and your agent should be using the app together. This creates new design challenges—agents can make a billion requests in three seconds, so you need approval flows, inboxes that summarize what happened, logs, and easy rollback.
10. Forward-deployed engineers are the new most essential role. The big model companies have teams of people managing their internal agents, and those teams aren’t going away. It’s different from traditional software building, and certain engineers love it. As models get better, this role will evolve—you’ll be managing more agents doing more things.
Automation is a lie. CLIs are over. The SaaSpocalypse is dumb.
A year ago @danshipper came on the podcast to predict where AI was heading. He was remarkably right—including the call that everyone was sleeping on Claude Code.
Dan has a unique lens into where things are going because his team at @every is possibly the most AI-pilled group of people in tech. I always learn a ton talking to Dan.
So I brought him back for round two. We'll score these in exactly a year:
🔸 Every company will have one “super-agent” in Slack.
🔸 Codex and Claude Code will become the new operating system for knowledge work.
🔸 The AI job apocalypse is not happening.
🔸 PMs and designers will thrive.
🔸 We will read way more AI-generated writing and we will like it.
🔸 "I would buy SaaS stocks right now."
Listen now 👇
youtube.com/watch?v=4D3hDmGh…
Dustin retweeted
Hearing this at dozens of midmarket companies too.
The heart of the problem is token waste. At least 90% of the tokens burned by Claude Cowork users are wasteful (i.e. same task could be performed -- and more accurately -- using better infrastructure and tooling.
Developers have a natural skillset advantage and a 1-2 year head start at practicing with this stuff, but Cowork took the rest of the corporate world by storm over the past few months, and the costs of getting this wrong are pretty bonkers, the inefficiency rate scaling multiplicatively with headcount.
The solution to token waste is a thoughtfully-built context graph, not a bunch of incumbent vendor APIs wrapped in wordy MCP calls.
A context graph shifts most of the reasoning burden from read time to write time. This makes a lot of the costs scale linearly with corpus.
New: @ServiceNow is the latest major public company to say it’s blown through its full year budget for AI coding tools from Anthropic in the first few months of 2026, just like @Uber CTO @praveenTweets said abt his company. “It’s a really hard problem,” CIO Kellie Romack said.
Dustin retweeted
We finally know why LLMs hallucinate. It's not the model. It's the geometry.
@OpenAI text-embedding-3-large: 91/3072 dimensions do real work.
@GeminiApp gemini-embedding-001: 80/3072 dimensions do real work.
~97% of your vector database is mathematically empty. Your RAG system is retrieving from noise.
@ashwingop and I present "The Geometry of Consolidation" - a proof that RAG compression has a hard floor no algorithm can beat, set by a single spectral number your embedding model cannot escape.
Every hallucination your RAG pipeline produces? This is why.
Paper + results: github.com/niashwin/geometry…
Dustin retweeted
This is big... Anthropic just announced a model so powerful they won't release it to the public out of fear over the damage it will cause 😨
Claude Mythos Preview found thousands of zero-day exploits in every major operating system and web browser...
The numbers are hard to believe:
> $50 to find a 27-year-old bug in OpenBSD, one of the most security-hardened operating systems ever built
> Under $1,000 to find AND build a fully working remote code execution exploit on FreeBSD that grants unauthenticated root access from anywhere on the internet
> Under $2,000 to chain together multiple Linux kernel vulnerabilities into a complete privilege escalation exploit
For context: these are the kinds of findings that previously required elite security researchers working for weeks.
Anthropic engineers with no formal security training asked Mythos to find exploits overnight. They woke up to working code the next morning.
The results were so impressive Anthropic assembled Apple, Google, Microsoft, Amazon, NVIDIA, and seven other organizations into Project Glasswing:
A $100M defensive coalition. They're not releasing this model publicly. Instead, they're racing to patch the world's infrastructure before models like this proliferate.
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.
anthropic.com/glasswing
Dustin retweeted
Welp, Anthropic's gonna stop the fun with OpenClaw.
Seems like a good excuse to open source what we've been cooking:
github.com/Endgame-Labs/goat…
Goated is an always-on personal AI assistant, built around Claude Code and Codex. It's minimal, performant, and piggybacks on the best harnesses in the world for long-running sessions.
Out of the box, Goated supports:
- Slack and Telegram chat interfaces
- Claude Code and Codex in both headless and TUI modes
- Long-running daemon operation with watchdog recovery
- Cron jobs and headless subagents
- CLI-driven credential management
- Session health checks, restart handling, and queueing
- A seeded private workspace/self repo with bundled note-taking tools and an extensible Cobra-based personal CLI
But it's just Claude Code (or Codex) under the hood, so your agent can really vibe anything you want -- and much better than OpenClaw.
It also doesn't leak memory.
Dustin retweeted
Software horror: litellm PyPI supply chain attack.
Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords.
LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm.
Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks.
Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages.
Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible.
