Something something in AI Safety
🤗 huggingface.co/abdelstark
Joined March 2012
- Tweets21.7K
- Following6.5K
- Followers30.9K
- Likes49.1K
Pinned Tweet
Can LLMs be PROVABLE computers?
Percepta showed that a transformer can BE a computer. Compiled weights, deterministic execution, 30k tokens/sec.
But nobody asked the obvious follow-up: how do you know it computed correctly?
So I built the verification layer. A STARK that proves it 👇
abdel retweeted
Dear @AbdelStark , the number one reason I get up in the morning eager for another day at StarkWare is the amazing team we have. Combination of gigantic brain, ambitious vision and values worth fighting for: integrity, human dignity and freedom.
A team is a collection of people, and you are one of the best examples of what StarkWare is:
Time and again you led the way — Bitcoin and freedom tech, Privacy, and most recently AI and the agentic economy and the safety challenges it poses.
Add to that you’re amazing charisma and technical power, single handedly willing full ecosystems and products to existence.
But above all it’s your big spirit and seeing how the values of human freedom and dignity are your beacons of light and North Star.
I feel more optimistic that AI won’t kill us all now that you’re working full time on saving us all.
Best wishes for your next chapter and I’m sure ZK STARKs and Starknet will play a role in AI safety too, so our paths will cross again.
Love respect and admiration, Sir!
Some personal news: after all those incredible years, I'm leaving StarkWare.
Joining StarkWare was the best professional decision I've ever made. It all started with @ukolodny, who convinced me with a single sentence that showed he understood exactly how I could have the biggest impact. What followed was the most epic, transformative and inspiring chapter of my 15+ year career. That's why it's really a difficult moment for me to leave, with such a deep emotional attachment to the people, to the project.
The people at StarkWare are out of this world. It's very easy to be the dumbest person in the room there, and honestly it's the best thing that can happen to you.
A few things I'll carry with me forever:
- The Exploration Team, a bold bet that @avihu28 and Uri trusted me with. We started many open-source projects that helped shape the Starknet ecosystem, showcase cool stuff that could be built with our tech, and some even became companies with their own journeys.
- The Starknet developer ecosystem, the fastest growing in the whole crypto industry two years in a row. But it was never about vanity metrics. It has a soul: we built alongside people, supported them, and watched them grow. And it was amazing to collaborate with such a great community of passionate people.
- The privilege of being at the forefront of StarkWare's bets on Bitcoin, privacy, freedom tech and AI.
So why leave now? Over the past year, AI safety took over my mind. I'm convinced it's one of the most important challenges humanity has ever faced (very likely the biggest imo), and that math, cryptography among other things have a big role to play in it. And also that we need much more people and resources to work on it. The call to make my humble contribution became too strong to ignore, also as a duty to my children and the world we leave them.
@EliBenSasson, it was an honor to ride with you. I believe the true impact of STARKs and verifiable computation is yet to be seen, for blockchains and far beyond. As you told me, our paths will likely cross again, I strongly believe it. There are so many people from StarkWare I would like to thank and name (luckily many of them are to busy shipping they don't even have an X account 😂)...
@bbrandtom my brother I love you (don't worry I will nurture the flame).
gigachad brains like @LiorGoldberg2 @VolokhIlia
You all know that I think the world of you, that working with you was such a pleasure, we had so many epic intense moments together.
I remain extremely bullish on Starknet and @StarkWare. The tech is years ahead, the StarkWare team is out of this world. The Starknet Foundation is full of great talented people too, and I have full confidence in @strjd87 ability to spearhead the vision and lead his team (full of great folks like @thedamooo @henrlihenrli @franklyteddy @omarespejel and others) towards success!
The Starknet ecosystem is still strong, and I am amazed to see friends like @MentorNotPseudo still on the frontline and shipping and battling constantly to help Starknet moving forward. I have great respect for the builders and I wish them all the best and great successes.
To the StarkWare team and the Starknet community: thank you for the trust, the energy and the friendships. If there's one team that can make StarkWare and @Starknet incredibly successful, it's this one. I'll be cheering from the front row.
Forever grateful ❤️
abdel retweeted
Should Starknet become an independent L1?
It's time for the community to have its say.
STRK stakers and delegaters can now participate in the first, non-binding governance vote on this potential transition 👇
governance.starknet.io/#/s:s…
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) ?
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) ?
ok... seem current models definitely improved in situational awareness...
xitter.cf/DKokotajlo/status/2110…
abdel retweeted
I tried building a word game using LLM hidden states. That sent me down a much deeper rabbit hole than expected.
From word2vec to attention to Jev, the whole detour explained from scratch
chqrles.me/en/words-in-conte…
Freedom tech is the most important component that can help us as individuals to take back our sovereignty. This applies for money as well as for intelligence.
Censorship resistant unstoppable money is what we need.
Similarly we need open source intelligence and avoid gatekeeping at all costs.
All of those things will require maths and cryptography among other things.
Privacy. Quantum resistance. Scalability. Censorship resistance.
Reminder that maths scale and that goodwill doesn’t.
abdel retweeted
1/ A powerful enough quantum computer could drain your wallet without stealing your seed phrase.
The fix: migrate to post-quantum cryptography in time.
We’re releasing a crypto-agility table and migration map, showing why some chains face a harder transition 🧵
quantum.starkware.co/network…
Your agent says “done”. I want receipts.
I fine-tuned @fastinoAI open-weight GLiNER2.5-Decide on my Mac to judge agent action logs.
910 training cases. 14.7 MiB LoRA published on Hugging Face.
Meet ExitReceipt: 158 test runs. Every prediction public.
Watch what changed ↓
The part I love about open weights: you can change a model, measure what changed, and share the actual adapter.
I published all 3 seeds. Two scored higher on test; I kept the one selected beforehand.
The release includes every prediction, including the mistakes.
Explore every case:
abdelstark.github.io/exitrec…
LoRA + model card:
huggingface.co/abdelstark/ex…
Code + training recipe:
github.com/AbdelStark/exitre…
What evidence would you require before letting an agent mark a task “done”?
abdel retweeted
Decision models should do more than pick a label.
To truly act, they need to extract evidence, understand relationships, produce structured records, and keep related decisions consistent. And they should do it fast.
That’s the new model we built with GLiNER2.5-Decide.
Like all GLiNER models, this one is highly responsive to fine-tuning. And now you can fine-tune GLiNER models, including GLiNER2.5-Decide, directly inside your coding agent with our new API.
To make building with GLiNER2.5-Decide as fast and simple as possible, we’re also offering hosted inference.
Our model goes beyond typed classification. It can return character-level spans, relations, and structured records, then apply implications, exclusions, cardinality limits, and ordinal bounds across related decisions.
And it does all of this very fast. On short, single-document requests:
CPU latency → 167 ms (48-vCPU Intel Xeon Platinum 8581C)
GPU latency → 38 to 47 ms (T4, L4, V100, A100)
Run it locally, host it, or fine-tune it. Your decision.
Download the weights on @huggingface: huggingface.co/fastino/GLiNE…
Submit your cool Jev project / experiment to the community awesome-typesafe repo 👇
370 stars. 56 contributors.
Jev community is hot.
github.com/AbdelStark/awesom…
500 stars. 100+ forks and contributors.
Growing nicely.
Awesome Jev.
github.com/AbdelStark/awesom…
