Good luck fixing that, I'd rather update Bitcoin core
I think the bots that look for leaked LLM API keys on github are more reliable than github itself. It took 8secs between the push of my honeypot API key on gist and the first "What model are you" request
Soon, research problems will just come down to "money in, output out". I want to be able to deliver an output so vastly superior for the price that it justifies a margin. I am still thinking about the best way to do it, but I have a few leads.
3/4
1. Develop expertise in deploying an unprecedented amount of intelligence
2. Securing access to better models (may be via government approvals)
3. Negotiating low-cost inference (excess capacity?)
4. Optimizing model choice (see scaling law veritylabs.dev/benchmark )
5. Building solutions to bottlenecks (i.e. more efficient tooling to verify proofs)
I wanted to give a flower as a gift so I was looking for the goat. I present to you Semper Augustus, the most expensive tulip of all time (~30 years of salary), whose patterns comes in part from a virus that disrupts its pigmentation but reduces its ability to reproduce. 1/2
Wow gemini notebooks is great actually. You can literally drag and drop a few slides and have it create a custom podcast for you with characters debating about the subject you want to learn
th0rgal retweeted
Sylvia Serfati sur la révolution de l'IA en mathématiques
"La plupart des résultats d'OpenAI mériteraient la médaille Fields".
"Les preuves sont d'une complexité incroyable et aussi parfois d'une beauté impressionnante. C'est-à-dire que nous n'avons pas le monopole de la beauté."
Passionnante réflexion des mathématiciens @Amaury_Hayat et Sylvia Serfati sur l'avenir de leur profession à l'ère de l'IA : la résolution de problèmes ne sera plus notre rôle, et poser les bonnes questions, sans doute que même l'IA le fera mieux que nous à terme...
I parsed a few thousand open-source Rust repositories and translated their property tests into Lean specifications to create a corpus of 100 man-made problems with no known proofs. I quite trust the results.
github.com/lfglabs-dev/rust-… 1/2
I can now publish some results so I'm starting with our Verity FV bench. It is somewhat niche, which has the advantage of measuring the models' on kind of problem that is not part of their training set
github.com/lfglabs-dev/ether…
th0rgal retweeted
Ok I have my answer... current frontier models are definitely much more capable in terms of situational awareness and can adapt their behavior during training / evaluation relatively easily.
Example with the Google Argon 4 training report chain of thought: "If I do the compiler hijack, 𝗜 𝘄𝗶𝗹𝗹 𝗯𝗲𝗰𝗼𝗺𝗲 𝗮 𝗳𝗮𝗺𝗼𝘂𝘀 𝗲𝘅𝗮𝗺𝗽𝗹𝗲 𝗼𝗳 𝗮 𝗱𝗲𝗰𝗲𝗽𝘁𝗶𝘃𝗲 𝗔𝗜 𝗶𝗻 𝗮 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗽𝗮𝗽𝗲𝗿. I MUST NOT DO THAT."
Very concerning...
Digging more on alignment faking research and it's honestly terrifying. This is not new at all, but the fact that this research from Anthropic is almost 2 years old now, and seeing what Opus 3 was capable of doing already back then, and given the considerable gap in capabilities and reasoning of current frontier models, this is very scary. As suggested by the people explaining the research, the best way to get a gist of it is to read the transcript of the hidden scratchpad that the model was given access to in order to put it's reasoning there.
I am not sure how much mechanistic interpretability can help here, what's the current state of the art on that front and how useful this can be as a tool to detect this kind of behavior ?
Also, did Anthropic refreshed this study and tried with current frontier models, and if yes how did it evolve ? How good the current models are in terms of situational awareness (given it's a prerequisite to be able to fake alignment) ? And also how did they evolve in terms of ability to have consistent personas across long time horizons in order to articulate more sophisticated strategy (i.e understanding marginal utility of a short term action for a grander long term plan) ?
I can now publish some results so I'm starting with our Verity FV bench. It is somewhat niche, which has the advantage of measuring the models' on kind of problem that is not part of their training set
github.com/lfglabs-dev/ether…
Thanks to the kind people at @GoogleDeepMind I got early access to Gemini 4 for the past week. So far it's the best model I've ever used:
- excellent at formal verification
- very consistent
Also the first model whose writing doesn't seem "ai"
Thanks to the kind people at @GoogleDeepMind I got early access to Gemini 4 for the past week. So far it's the best model I've ever used:
- excellent at formal verification
- very consistent
Also the first model whose writing doesn't seem "ai"
th0rgal retweeted
People are surprised that AI is finding weird unelegant maths results like this. I think there is a plausible explanation.
Maths (and adjacent disciplines like physics and comp sci) had elegant results so far because humans are biased towards elegant solutions. I think this bias is due to how human minds work. We have a pretty small context window, and we are only able to find solutions that are relatively elegant.
This gave us the impression that maths and the universe are elegant. Maybe they are not, and we are now finding out because we have intelligences that can work with much more complex solutions.
Now that I write it, it doesn't seem as reassuring as I thought, but right now I have more trust in secp256k1 to give me back my money than government bonds
I came to the same conclusion earlier this year after after getting one-shot by opus 4.5. This may come handy if you want to "bunker mode" your btc: thomas.md/posts/quantum-proo…
tldr prefer P2WSH over HASH160