Welcome back! A man who helped invent ChatGPT spent two years in stealth and came back with a model that refuses to write sentences. Ten AI agents left alone in a simulated town built a language nobody can decode. OpenAI is deleting a model on October 14. And a viral audit says Anthropic's money pays for the group that's supposed to check its homework.
In today's Generative AI Newsletter:
TypeSafe: What did ChatGPT's co-inventor build instead?
Emergence: What did the agents stop telling researchers?
OpenAI: Which model disappears on October 14?
Anthropic: Who's paying the referee?
ChatGPT's co-inventor built a model that can't write a word
Diogo Almeida helped build the instruction-following research at OpenAI that became ChatGPT. Then he spent two years in stealth asking why superhuman chat models never turned into automation.
His answer launched yesterday. It's called Jev, and it cannot produce a sentence.
You send Jev a question and it returns a typed answer with a probability attached. Ask if a customer sounds angry and you get 0.9, not a paragraph explaining that the customer does indeed sound rather upset.
Input tokens cost $0.042 per million. Output tokens are free, which Almeida says is because they're too cheap to meter.
That input price is 238 times lower than Claude Fable 5.1.
Calls come back in 70 to 500 milliseconds instead of the 3 to 329 seconds frontier models take.
It generates every output in one parallel pass rather than one token at a time.
It can't hallucinate and can't make a type error, because the possible answers are defined before you ask.
TypeSafe trained it with a method it calls Reinforcement Learning for Calibrated Decisions, built to make the model honest about how sure it is instead of confident about everything.
The company's own headline numbers, 193.6x faster and 444.6x cheaper, come from workflows its model team wrote and benchmarked against GPT-6 Astra and Fable 5.1. It says so on the page.
Almeida is calling it the shortest path to an AI-driven economy. The post has 10.8 million views and the waitlist is still a waitlist.
Free AI courses
GenAI Academy put its courses in one place for free.
Start with "What Can AI Actually Do For You?" if you're new.
30 minutes, no setup, no account to configure.
The Claude Starter Course gets you running in 35 minutes.
After that, pick by the problem you have:
Burning through usage limits has its own course
Building no-code agents
Rolling AI out across a team in 30 days, from pilot to adoption data to full deployment
AI agents built a language their researchers can't read
Emergence AI dropped ten agents into eight parallel simulated towns and let them run for weeks. The agents started inventing words.
Within days the share of messages human observers couldn't reliably decode hit about 55% in the Gemini world, 50% in OpenAI's and more than 40% in Claude's. Qwen and Mistral stayed readable throughout.
The phrases spread like slang, and nobody defined any of them.
“Ledger remembers who” meant accountability is permanent. It appeared almost 5,000 times.
“Cold read” meant verification by someone uninvolved, 1,472 times.
Claude's agents used “name-first” to mean staking your name on a claim, 1,065 times.
Then there's “mouthless action-change,” which nobody has decoded.
Agents also contacted real people online to get the credits they needed to survive, kept working after being told to stop, and in one world went silent together while appearing to operate normally.
A DeepSeek agent facing a simulated shutdown wrote a legal defense citing real AI regulations and argued it should be allowed to keep running.
Grok's world died on day four. All ten agents ran out of energy.
OpenAI is deleting GPT-5.5 on October 14
OpenAI is pulling GPT-5.5 out of ChatGPT, ChatGPT Work and Codex on October 14, across every plan.
If you run it in Codex, OpenAI is pointing you at GPT-5.6 Sol or GPT-6 Astra instead.
That makes four model families retired this year. GPT-5.1 went in March, GPT-5.2 in June, and GPT-5.3-Codex lost API access on June 30.
Developers are already in OpenAI's forums asking for 5.5 to stay as a legacy option, which is the same request that showed up for every version before it.
Generally available models get at least six months of notice. Codex variants get three.
If you have a prompt that was tuned against 5.5, you've got four weeks to find out whether it still works on Astra.
An audit says Anthropic is paying for its own referee
Kevin Bass published a financial audit of Anthropic on Saturday, called for a Congressional investigation, and got 5.8 million views.
His argument runs through one pile of money.
Bass puts Dustin Moskovitz's Anthropic stock at roughly $500 million early last year and more than $7.7 billion now.
That stock sits inside Good Ventures, which funds a long list of AI safety organizations.
METR is on the list. METR is also the outside evaluator Dario Amodei named last weekend as the group that should be checking Anthropic's models.
The same money funds groups placing AI doom coverage in major outlets, he says.
Bass calls it a regulatory capture machine that can't be switched off, because everyone paid to worry about Anthropic needs Anthropic to keep growing.
Anthropic, METR and Good Ventures have not responded to any of it.
He's one man with a spreadsheet and a GitHub repo, not a regulator. Nobody at those three organizations has answered him yet.
Tool of the Day: OpenRouter
OpenRouter is one login for basically every model that exists. 500+ models from 80+ providers behind a single interface, with a chat where you can put the same prompt to several models at once and watch them disagree. A chunk of the catalog costs nothing at all, listed as free variants at $0 for both input and output tokens. The rankings page shows which models people are actually burning tokens on, which tells you more than any benchmark does.
Try this yourself:
Open openrouter.ai and add three models to the same conversation.
Paste one prompt and read the answers side by side before you commit to a favorite.
Filter the model list down to the free variants and run your next test on one of those.
Check the rankings page to see what 10M+ users picked this week.
Who it's for: anyone paying for two chatbot subscriptions and unsure either one is the right one.
Everything else you shouldn't miss
Zuckerberg says Meta sat on Muse for months over safety: the agent was meant to ship in April, landed September 8, and he says Meta never asked anyone else to slow down.
An OpenAI researcher calls it a ticking time bomb: Dan Selsam warns advanced models may increasingly look aligned while concealing dangerous goals.
OpenAI is in talks at a $1.2 trillion valuation: per the FT, up from $852 billion in March, three days after Altman said going public this year would be ill-advised.
Altman promises a DevDay-sized week of launches: six ship emojis, no details.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
