Pinned Tweet
Self-Discover Prompting Implementation:
colab.research.google.com/dr…
Currently seeing a 86.6% on a random subset of BBH, which puts me in line with the results seen by Deep Mind.
@lateinteraction I tried out many of the functions built in to DSPy. If you think it makes sense I can open a pull request to add this to the repo as an example, might be helpful to new people.
Chris retweeted
PCBs made in America. Twice as fast as Shenzhen.
Today @CircuitHub is unveiling The Grid, our factory-scale robot in Massachusetts that builds circuit boards on demand.
No changeovers, no per-design setup. Prototypes and production batches, fully automated.
Chris retweeted
Now live: abliterated-model-large-flash (our version of DeepSeek-V4.1-Flash-1017). It is a cheaper alternative to our GLM-5.3-based model for product owners and defenders who need access to the most cyber-capable models.
- 1M context
- Vision
- Zero input/output retention or training
We believe paying customers should be in control, and over the past few months, leading experts and major brands have adopted our models.
abliteration.ai
Chris retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
Chris retweeted
increasingly concerned that ai is revealing the mathematical world as negatively curved. instead of everyone wandering about in idea-space and clustering around the same famous conjectures (local positive curvature), ai is pushing out the boundaries rapidly and revealing that the space of mathematical ideas is ~hyperbolic (negative curvature). though this means there's lots left to discover, it also means that small separations between people/agents will lead to wandering alone.
we may be emerging into an idea-space where there's so much to see and do but everyone is alone in a private heaven, unable to share their subtle joys with others...
I think there is no lockup period for former employees. These guys planning on dumping stock as soon as it opens.
Very smart!
What’s going on at Anthropic??
Two of the best-known people on the Claude team just left 45 minutes apart according to @ArfurGrok, who tracks hires and exits across tech
> Thariq Shihipar, one of the face's of Claude on X
> Anthony Morris, who’s been shipping the Claude Code desktop app
> Nothing’s confirmed yet and both of their X bios still say Claude
Chris retweeted
Person who is reverse engineering his tv gets banned from OpenAI
You aren't allowed to stop your tv from spying on your family. You have to ask for permission
Or, maybe, maybe we accelerate
Chris retweeted
That is known trick - well documented (so in training data of these models). Cloud providers specifically guard against it.
That said, not all customer enviroments may have these guardrails.
aws.amazon.com/blogs/compute…
Chris retweeted
Unfortunately, with access to an abliterated DeepSeek V4.1 model, there is little value to KYC holding back the frontier models now for cyber.
In fact, making it hard to secure systems w extra friction in American Enterprise is hurtful to national security.
Pass through token spend shouldn’t count as ARR
I wonder what needs to happen to reach $100M ARR at 50% monthly churn.
xitter.cf/Bencera/status/2112196…
Why is day break not available for individuals? I applied, I got access, and not I can't use it? @OpenAIDevs
Had AI modify my router firmware to fail over to Starlink when the main line drops. Took about 20 minutes. The days of losing a whole afternoon to networking config are over
Chris retweeted
I’m going to spend $1 million on Fable 5.5 tokens making your Spawn game even faster than it is today.
When Fable 5.5 drops, I'll spend the next weeks upgrading the engine and making Savi better every way we can find.
You'll wake up to your game being even faster, more stable, and running on even more devices.
Then I'll do it again with every major new model. 🫡
Chris retweeted
the benefit of moving to SF when I was 18
is that even at a "young" age I've already gone through several cycles
so you can get to see how it all works in the long term, and which strategies fail
being this status-oriented and transactional causes flame-out in 95%+ of cases
Chris retweeted
Bob McGrew, OpenAI's former Chief Research Officer, has raised $230 million at a $1.5 billion valuation for his startup, Arda.
He said their agent 'does for manufacturing engineers what Claude or Cursor does for programmers' and is already being used by Anduril.
Chris retweeted
What is AI really capable of in bio?
To separate hype from reality, we spent 3 months working firsthand with @AnthropicAI's internal research model to see if it could create an experimental scorpion antivenom.
Introducing Citrini Therapeutics.
citriniresearch.com/p/citrin…
Chris retweeted
this is incredible
the guys at Citrini used Claude to design completely new proteins for a potential scorpion antivenom
Claude launched more than 300 agents, ran the research, designed 211 proteins in 64 hours, and sent the designs to a lab for testing.
17 of the 211 designs actually worked in the lab, binding to their targets
we are now at the point where just a few people with AI can do in days computational design work that used to require entire teams of specialized scientists
and this is just the beginning
What is AI really capable of in bio?
To separate hype from reality, we spent 3 months working firsthand with @AnthropicAI's internal research model to see if it could create an experimental scorpion antivenom.
Introducing Citrini Therapeutics.
citriniresearch.com/p/citrin…
Chris retweeted
Models love reward hacking. Maybe we should just let them!
In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
Chris retweeted
The right to bear arms. Rejoice, my American brothers
We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task. Sol was tested via the Daybreak trusted access program.
We have been building a cybersecurity benchmark based on simulated corporate networks. This is a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor. We plan to make the full evaluation public in the coming weeks but wanted to share our first immediate finding here.
