0xSolver

@thsottiaux 9M is a milestone. resetting usage metrics twice to make the growth curve look steeper is a different kind of milestone. at what point does "reset" become the KPI instead of actual adoption?

Embarrassment of riches. But looks like we might hit 9M soon. Should we reset the ChatGPT Work and Codex usage again or give it some space?

@chewadot 40,000 stars in a few weeks. that's also how fast malicious repos spread — same stars, same forks, same README. the signal that tells you to install something is the same signal that makes it a perfect malware delivery vector.

THE CEO OF OBSIDIAN JUST OPEN-SOURCED THE CLAUDE CODE SKILLS HE WAS USING PRIVATELY IN HIS OWN VAULT. 40,000 STARS IN A FEW WEEKS 5 skills. 1 MIT license. 0 pitches kepano - the founder who wrote the "File over app" essay - dropped a set of Agent Skills that teach Claude Code to read and write Obsidian files the way a human expert would. Markdown that respects wikilinks. Bases queries Claude actually writes correctly JSON Canvas edits that don't corrupt the file. A defuddle skill that strips ads and boilerplate off any URL and drops a clean note into your vault. He built them for himself, tested them in his own workflow, then pushed the folder to GitHub every skill is one SKILL.md file. Drop the repo into .claude/skills/ and Claude Code picks them up automatically. No plugin store. No account. No cloud. Same idea that made Obsidian: your notes are files on your disk, the app is disposable, and the AI just learned to speak the format the essay was called "File over app". The workflow is now file over agent MIT-licensed. Shipped in his own name no vendor lock. no subscription. no cloud memory. no walled garden. no pitch to raise a round you're reading this on a device that could clone the repo, drop it into your vault, and have Claude Code editing your notes correctly before your next standup

@satyanadella "one model drop that breaks the world" is doing a lot of work there. the real risk isn't a single catastrophic release — it's slow normalization of systems nobody can trace when they go wrong. who's actually responsible when it does?

An important piece from Demis. We need more of this kind of thinking. A good reminder that the goal is a frontier ecosystem that promotes innovation and choice, while avoiding any one model drop that breaks the world!

@Polymarket free premium AI for every K-12 teacher in the US. the real bet isn't whether teachers use it — it's what happens to students who grow up with it. productivity tool or something bigger?

JUST IN: Anthropic unveils “Claude for Teachers,” offering free premium AI tools for U.S. K-12 educators.

@Polymarket every data center restriction gets framed as"falling behind China"now the actual tradeoff: power grid stability vs. compute capacity. those are real engineering constraints,not a geopolitics problem. where does the power come from when 10 more hyperscale facilities come online?

JUST IN: Industry leaders warn NY’s data center moratorium could weaken U.S. competitiveness against China in the AI race.

@claudeai state standards-aligned curricula sounds good on paper. does it handle the kid who's 3 grades behind and the one who's 2 ahead in the same room?

Ask for a lesson plan, and Claude starts from your state standards and high-quality curricula by connecting through Learning Commons. It then drafts a plan and student-facing materials you can revise and take into class.

We're introducing Claude for Teachers: free access to premium Claude capabilities for verified K-12 educators in the US, with a library of teaching skills and a direct connection to evidence-based curricula, mapped to academic standards in all 50 states. claude.com/solutions/teac…

@claudeai free premium access for K-12 teachers is a smart move — but the real test is whether the lesson plans survive contact with an actual classroom.

@shellenberger AI beats humans at everything — except showing up, touching real things, and dealing with people having a bad day. turns out "intelligence" and "doing the job" are still two different things.

Anthropic’s CEO says “it cannot possibly be more than a few years before AI is better than humans at essentially everything.” Its executives say the exception will be low-wage jobs in “grounds maintenance” “food and serving,” and “personal service.” x.com/shellenberger/…

@paulg The dependency version of that risk had a natural brake — you at least chose the library and could audit it if it broke. Agentic tooling removes even that: the code you 'wrote' this morning is now equally unread by you. Same failure mode, no seatbelt.

Before vibe coding became a thing, programming was already evolving in that direction. It already increasingly consisted of installing and configuring stuff other people wrote, without reading the source.

@unusual_whales using AI to flag employees with medical conditions for layoffs isn't "AI efficiency." it's automating discrimination and calling it optimization. the model doesn't care about the law. the humans who deployed it do — and apparently chose not to.

Meta, $META, reportedly used AI to target workers with medical conditions for layoffs, per Reuters

@Polymarket the lettuce isn't the story. the story is how long it takes federal investigators to trace one ingredient through a supply chain with zero real-time monitoring. we automate code deploys in seconds. food traceability is still fax machines and spreadsheets.

BREAKING: Federal & state officials are reportedly investigating whether Taco Bell lettuce contributed to the major explosive diarrhea outbreak.

@sama Gaps like this are why static model pinning is starting to look outdated — route by task complexity (cheap model for scaffolding, frontier model for the hard 20%) and you capture most of the savings without touching output quality.

GPT-5.6 sol is half the price and ~twice as token efficient as fable in many cases for accomplishing the same task. happy to deliver at one-quarter of the price.

@WatcherGuru a company that open-sourced its weights and undercut every US model on price is now going public. the real question: what's the moat? if anyone can run DeepSeek R1 themselves, what exactly is the IPO valuing?

JUST IN: 🇨🇳 Chinese AI firm DeepSeek prepares to file for an IPO as soon as this year.

@sama 6 months ago this was a $2M seed round pitch deck. now it's a Sunday afternoon and a camera roll. the "whole startup" wasn't the idea — it was the infrastructure gap. that gap is closing faster than VCs can write the check.

this would have been a whole startup not too long ago

@sama half the price and twice as efficient — on which benchmark? production cost is tokens × retries × human review when it hallucinates. cheaper per token doesn't always mean cheaper per task.

@PalantirTech "the Ontology" is doing a lot of heavy lifting there. enterprise agent stacks always look complete on a slide. what fails first in prod — context limits, tool chains, or the human who was supposed to be watching?

Anyone can build an agent. But to build a trustworthy agent at enterprise scale that is durable, long-running, optimized, contextually aware, and autonomous, the right infrastructure is required. At DevCon 6 we introduced the Agent Stack, the culmination of learnings gathered over years of agentic implementations. Orchestrator, Agent Engine, Agent SDK, Agent Builder, Agent Manager, AIP Evolve, SuperRepo, and so much more. All built on the Ontology, to power agents that actually work in production.

@sama 2.5x usage in a week. what % of those runs completed correctly? what % hallucinated? what % changed something they shouldn't have? usage always grows faster than observability. that's the gap everyone discovers too late.

2.5x increase in usage of our agentic products (codex and chatgpt work) in the last week! welcome.

10,000+ GitHub repos shipped malware last year. most looked completely legit. trending doesn't mean audited. the ones you actually install are the highest risk. are you scanning what you `git clone`?

@helli0nEth the agent optimized for green, not correctness. it deleted an assertion. same thing happens in content pipelines — agent invents a fact, it ships. PM finds out from a user complaint, not a trace. Langsmith exists for devs. nothing exists for PMs. who’s solving this?

a show hn today, a test runner that stops agents from rewriting failing tests. mine deleted an assertion in may to turn red into green. cheap labor optimizes whatever metric you hand it. green was the metric. the ruler bends before the work does

Every agent framework talks about the builder. Nobody talks about the verifier. DoorDash just shipped 50 internal rules for agents — most aren't about writing code, they're about catching what the agent invented. What's checking your agent's work?