·
Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
PLUS: AI agents that drain your brain
Nvidia is backstopping $250 billion in financing for OpenAI to lease a 10 gigawatt datacenter in Ohio. Half a trillion total.
Open weights matter because they let you verify claims and avoid sharing data with closed labs. Six new models dropped this week alone.
Nvidia started with the wrong algorithm, bought three textbooks at Fry's, and invented most breakthroughs in modern computing. Resilience matters most.
Hermes Agent now scales to near infinite MCP tools with high accuracy using progressive disclosure. Tools behave like skills behind search.
AI model guides are outdated within 48 hours now. Opus 5 and Codex voice both launched Friday.
Sam Altman asked ChatGPT to plan a trip, build a site, coordinate friends, and make reservations. It just worked.
Claude's Excel extension researched for 37 minutes and created an in-depth Nvidia analysis from one prompt. Opus 5 is wild.
Smarter models don't reduce the need for applied AI layers that bridge intelligence with real workflows. More capability demands more integration.
Jason at OpenAI uses Codex to turn Slack messages into speaking skills and check into flights automatically. The handbook author shows how.
Remote companies have an AI advantage because work exhaust is documented. Zapier's institutional knowledge supercharges new hires instantly.
Agentic workflows need loops and graphs where output decides next steps. But you stay in it. You're not the operator.
The biggest bottleneck to scaling autonomous mining is physically retrofitting decades-old equipment and calibrating sensors on site. Very different from software.
Smalltalk wanted everything minimal and elegant. HyperCard was just a friendly drawing app until you made a game. Smart programmers underestimate that.
Chinese media says OpenAI's business model rests on scarcity but now you can get frontier capability for free. The distillation accusations blend concern with politics.
Anthropic shipped every Claude model from Claude 2 to Mythos with just five product engineers. Evals are the new PRDs.
Dictate a brief, let Notion trace code and write the spec, hand it to an agent. Review swarms catch bugs while you're gone.
First you accepted every line. Then you let it rip. Now the agent stops at checkpoints to explain what it did and what's next.
Most people spend three days picking a model and zero days deciding what the agent owns. Answer those first.
The MCP final spec drops in two days. You're downstream of decisions you had zero input on and find out with everyone else.
Revenue is stabilizing but renting an AI CEO to small businesses has a cleaner exit than optimizing the marketplace.
ChatGPT Voice lets you work anywhere through AirPods and control your desktop. Four hour hike became four hours of work.
Running a prompt to open Chrome, read Twitter, and summarize what's interesting is just normal now. And it works.
Claude Code includes 1M token context windows in your subscription but charges for fast mode. Codex does the opposite.
Startup School ran hundreds of small-group sessions with YC partners after feedback that people wanted less big talks.
Selling courses and PDFs starts going stale day one. A community whose product is software beats every content seller. $250K from Claude Code.
AI agents running somewhere use up mental energy even when they're useful. The mere existence causes multitask mode.
AI can write new episodes that capture humor and rhythms better than actual writers. Netflix premium could generate shows on demand.
Opus has the GPT problem of adding 100 lines of fixes when it notices another issue. Still need Fable to trim slop.
AI replacing all jobs means we'll be fabulously wealthy. A tsunami that creates wealth, not destroys it. Rich man's problem.
Games are the software humans most want to use themselves versus delegate to agents.
People without AI psychosis don't worry about tokens. They worry about trusting ChatGPT with Gmail and Google Workspace.
Replacing customer support with AI is backwards. There's a real premium for excellence when everything around you is pretty good.
This is why we need competition and open weights. Imagine only Anthropic as moral arbiter of acceptable speech.
Anthropic is saving Fable 5.1 for OpenAI's next release. They'll keep crossing swords. Fable-class moved to credits preemptively.
Frontend slop feels much easier to fix than backend slop. User visible problems but somehow more acceptable.
Kimi is scaring the shit out of OpenAI and Anthropic. This is a good thing.