We can finally talk about it:
We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company.
We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.
Let's cut through the untrue sensationalist reporting.
- OpenAI is NOT going to come after random people and demand that they share profits from discoveries made with the aid of ChatGPT.
- OpenAI is NOT going to "force" anyone to hand over their profits.
Here's what's actually going on: OpenAI will approach enterprise customers and offer them to sign a deal that will include a revenue share with OpenAI from products made with the aid of OpenAI's models. It's up to the enterprise customer - a sophisticated party, represented by competent outside counsel - to agree or disagree to pursue this arrangement.
We have a recent quote from Sarah Friar on this:
"I like [the idea of] licensing models to really align [the customer's and OpenAI's interests]. Let's say in drug discovery, if we licensed our technology, you have a breakthrough. That drug takes off and we get a licensed portion of all its sales."
See also the screenshot below from OpenAI's recent blog, which talks about "licensing, IP-based agreements and outcome-based pricing" that would let OpenAI share in the value created by its models.
@karpathy Added! In the next version of Claude Code, you can use the CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS env var.
eg. "CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS=1234567 claude"
You can also add this to the "env" section in your settings.json
@mayankja1n I still have mine! Thalmic’s problem wasn’t the product, it was great. The issue was it was too technical for mass-market, no clear killer use case, the community never grew, and it wasn’t cheap. Maybe just a bit ahead of its time.
Today, we are taking the first step to transform @raycastapp into a Natural Language Launcher with the beta release of AI Extensions for our Pro subscribers. Simply type what you want to do, and Raycast AI will take care of it for you.
Let me show you some examples...
1/8
Si DeepSeek no puede explicar la masacre de Tiananmén entonces es una herramienta política en un nuevo formato y no sirve como fuente de consulta, al igual que TikTok y las redes sociales no reflejan la realidad sino el resultado de algoritmos
La foto no reemplaza al paisaje
OpenAI researcher asking the right questions
🧠 Fun fact 🧠
To prove a point, Eliezer - a mere human, role-playing as an AI in a box - once convinced two people to let him out... despite them being adamant they would NEVER do so?
Eliezer never explained how.
Anyway, sigh, we aren't even using boxes, we just hooked them up to the internet immediately.
"So how did the AIs escape the box in the end after all the precautions?"
"Box?"
@simonw Currently, the API model is extremely slow and does not give any feedback or reasoning. Even if you try to use an old prompt where you request a scratchpad, it constantly refuses to provide any reasoning for its response.
At the height of One Million Checkboxes's popularity I thought I'd been hacked. A few hours later I was tearing up, extraordinarily proud of some brilliant teens.
A thread about my favorite story from running OMCB....
@pimoroni Are WD drives good for NVMe Base Duo, or is this just an outdated list? It appears in the supported disks of the Duo, but from the base with one slot, it appears in the list of disks to avoid.
https://t.co/7oYBwIvBej
https://t.co/QShoUQKdCg