LAGRANGE (@lagrangedev) header
LAGRANGE (@lagrangedev) profile picture

LAGRANGE

@lagrangedev

DeepProve: a cryptographic receipt for every AI inference. now open source. Halo: autonomous coordination engine for the toughest conditions. @LagrangeFndn $LA

 Los Angeles
lagrange.dev
Joined May 2022
  • Tweets3.8K
  • Following378
  • Followers93.9K
905Media
Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)Photo by LAGRANGE (@lagrangedev)
The Frontier Red Team at @AnthropicAI just made the case that you should be able to check which AI model answered you, without ever seeing its weights. Their research names DeepProve “the current peer-reviewed state of the art.”
Image from LAGRANGE (@lagrangedev)
3
8
5
29
4,074
LAGRANGE retweeted
Lauris (@lzminsky) profile picture
Lauris
@lzminskyNov 6
I've seen @lagrangedev and StarkNet being mentioned here, but only a few know the connection between this article by @AnthropicAI and Rari Capital, if anyone here still remembers that. Significant motion on part of 2020/2021 DeFi devs, a cohort amazingly abundant in s++ talent nitter.gq/AnthropicAI/status/211…
Anthropic (@AnthropicAI) profile pictureAnthropic
@AnthropicAINov 3
New Anthropic research: progress on provable inference for large language models. Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers. For our full write-up, see here: anthropic.com/research/prova… Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP… We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
Progress on provable inference for large language models

Progress on provable inference for large language models

Sharing progress on a powerful type of cryptographic proof known as a zkSNARK, which is uniquely fit for provable inference.

anthropic.com
8
2
26
2,122
클로드(Claude) 개발사 앤트로픽(Anthropic)에서 라그랑주 딥프로브(DeepProve)를 언급했습니다!!! @AnthropicAI가 공개한 AI 추론 검증 연구에 @lagrangedev의 DeepProve가 소개됐습니다. 선행 연구를 다루는 부분에서 '동료 심사를 거친 현재 최고 수준의 기술 중 하나'라고 평가한 점이 눈에 들어오는데요. 어떤 기술인지 쉽게 풀어보면 이렇습니다. "지금 받은 답변, 정말 내가 선택한 AI 모델이 만든 게 맞을까?" AI 서비스를 이용할 때 실제로 어떤 모델이 답변했는지 사용자가 직접 확인하기는 어렵습니다. 딥프로브(DeepProve)는 바로 이 부분을 다룹니다. 답변 전체가 지정된 모델의 계산 결과인지 영지식증명으로 검증하는 기술입니다. 앤트로픽(Anthropic)은 이번 연구에서 딥프로브(DeepProve)의 검증 방식과 처리 성능도 함께 소개했습니다. AI가 점점 더 많은 일을 맡게 되는 만큼, 어떤 모델이 그 결과를 만들었는지 확인하는 기술도 중요해지는 것 같습니다. 라그랑주 앰배서더로서 정말 반가운 소식이네요. 꾸준히 쌓아온 연구가 이런 곳에서 언급되는 모습을 보니 앞으로의 발전도 더 기대하게 됩니다. 🔗 앤트로픽(Anthropic) 공식 연구 : anthropic.com/research/prova…) 🔗 딥프로브(DeepProve) 논문 : eprint.iacr.org/2026/1112
🤖 Made with AI
Image from 리도스 Lithos (@Lithos__x)
LAGRANGE (@lagrangedev) profile pictureLAGRANGE
@lagrangedevNov 3
Proud that @AnthropicAI’s research recognizes @lagrangedev's DeepProve as the current peer-reviewed state of the art. Every AI answer should come with a proof of which model gave it. Trust is moving from “take our word for it” to “check it yourself,” and today a frontier lab started agreeing.
69
43
2
84
1,168
A player jammed our swarm 389 times over Los Angeles, for more than ten minutes. It lost no drones and finished all eight sites 30 seconds after the jamming stopped. How long can you hold it? holdtheswarm.lagrange.dev
2
4
24
1,712
For most of computing, checking someone's work meant doing it again. You cannot do that with an AI inference. You do not have the weights. You do not have the input data. You do not have their hardware. Re-running is not expensive, it is unavailable. That is why a proof is not a convenience. It is the only check left. Anthropic's Frontier Red Team made the same argument this week.
2
4
9
1,851
No.3, who made the record. nitter.gq/lagrangedev/status/210…
LAGRANGE (@lagrangedev) profile pictureLAGRANGE
@lagrangedevOct 7
Every trust mechanism in software comes down to one question. Who made the record. A log is made by the system, afterwards, and it says what the system decided to write. An audit is made by a person, later, out of those logs. A proof is made by the computation, while it runs. Nobody writes it. Nobody can edit it. That is the whole difference.
1
3
291
Prove you're over 18 without showing your birthday. Prove the answer is right without showing the work. DeepProve does that for AI inference, and @AnthropicAI's research team just named it the current peer-reviewed state of the art in the field. @chbpap explains the idea.
LAGRANGE (@lagrangedev) profile pictureLAGRANGE
@lagrangedevOct 14
A model gets retrained on a Tuesday and every proof about it has to start over. @chbpap, our chief scientist, walks through the Eurocrypt 2026 paper on proofs that update instead.
1
8
19
1,951
Proud that @AnthropicAI’s research recognizes @lagrangedev's DeepProve as the current peer-reviewed state of the art. Every AI answer should come with a proof of which model gave it. Trust is moving from “take our word for it” to “check it yourself,” and today a frontier lab started agreeing.
Anthropic (@AnthropicAI) profile pictureAnthropic
@AnthropicAINov 3
New Anthropic research: progress on provable inference for large language models. Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so. These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers. For our full write-up, see here: anthropic.com/research/prova… Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP… We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
Progress on provable inference for large language models

Progress on provable inference for large language models

Sharing progress on a powerful type of cryptographic proof known as a zkSNARK, which is uniquely fit for provable inference.

anthropic.com
2
6
2
18
2,947
Big congrats to @chbpap and the team behind DeepProve. Read the paper: eprint.iacr.org/2026/1112
5
365
Incredible to see @AnthropicAI refer to @lagrangedev’s DeepProve as the “the current peer-reviewed state-of-the-art” for verifiable inference. The frontier of ZK, verifiable and private inference is moving faster than ever!
Image from Ismael Hishon-Rezaizadeh ⭕ (@Ismael_H_R)
8
15
95
4,603
The why is the best part: catching a swapped model, a tampered server, or a fake “our best model checked this.” Their post, with the code and proofs, all open source: anthropic.com/research/prova…
Progress on provable inference for large language models

Progress on provable inference for large language models

Sharing progress on a powerful type of cryptographic proof known as a zkSNARK, which is uniquely fit for provable inference.

anthropic.com
3
291
A drone that flies on someone else's network flies at someone else's settings. In October, one of those settings turned out to be a maximum speed of 100 mph.
Lagrange header. Headline: Ukraine's drones hit a speed limit in the sky. It was 100 miles an hour. A speedometer needle sits at 160 km/h, where the satellite link drops, far short of the 370 km/h jet-powered Shahed marked at the top of the dial.

ALT Lagrange header. Headline: Ukraine's drones hit a speed limit in the sky. It was 100 miles an hour. A speedometer needle sits at 160 km/h, where the satellite link drops, far short of the 370 km/h jet-powered Shahed marked at the top of the dial.

1
3
10
1,802
Cheap satellite link, or a network the drones build themselves. What each one costs: lagrange.dev/blog/ukraines-d…
Ukraine's Drones Hit a 100 MPH Speed Limit | Lagrange

Ukraine's Drones Hit a 100 MPH Speed Limit | Lagrange

For ten days in October, Ukrainian drones lost their satellite link whenever they flew too fast. The limit matched a number printed on Starlink’s own help page. Russia, cut off from Starlink months...

lagrange.dev
1
3
400
What’s the scariest thing a machine has ever told you? We’ll go first: “Task completed successfully.”
2
4
12
2,258
Somebody tells you the model was safe. The drone stayed in its lane. The agent bought what you asked for. Maybe. The question is not whether they are lying, because mostly they are not. The question is what happens if they are wrong, and how anyone outside the building would ever find out. Everything we build is an answer to that question.
1
5
10
1,995
The second one, from earlier this month. nitter.gq/lagrangedev/status/210…
LAGRANGE (@lagrangedev) profile pictureLAGRANGE
@lagrangedevOct 6
Every autonomous system reaches a moment where a person has to put their name under what it did. Not a dashboard. A signature. Someone accountable, saying this happened and here is how I know. We used to manage that by trusting whoever built it. That stopped working when the model started deciding before anyone could look, and the swarm started deciding a thousand times before anyone looked up. How do I know? Everything we build is an answer to that question.
4
445
Someone in the group chat has not replied for an hour. You cannot tell if they are busy or if their phone is dead, so you wait. Episode 10, the last one: a drone swarm has that exact problem and no way to ask. Under a minute, zero jargon.
2
4
1
12
2,269
Episode 9, if you missed it: more drones is not the answer, and the reason is how much each one is allowed to say. nitter.gq/lagrangedev/status/211…
LAGRANGE (@lagrangedev) profile pictureLAGRANGE
@lagrangedevOct 15
No meeting ever got faster when more people joined. Add one person and you add every conversation they are now in: six people make 15, fifty make 1200. Episode 9: why a swarm scales when every drone knows less.
1
1
4
398
The series, from the start: 1. What does jamming actually do? nitter.gq/lagrangedev/status/208… 2. Why is GPS so easy to deny? nitter.gq/lagrangedev/status/209… 3. Why is one drone easy, but fifty drones hard? nitter.gq/lagrangedev/status/209… 4. Can one shot take down a whole swarm? nitter.gq/lagrangedev/status/209… 5. What happens when a swarm loses half its drones? nitter.gq/lagrangedev/status/210… 6. What's worse than jamming? nitter.gq/lagrangedev/status/210… 7. Why can't one pilot fly fifty drones? nitter.gq/lagrangedev/status/210… 8. What does autonomous actually mean? nitter.gq/lagrangedev/status/210… 9. Why isn't more drones the answer? nitter.gq/lagrangedev/status/211…
LAGRANGE (@lagrangedev) profile pictureLAGRANGE
@lagrangedevSep 10
Take fifty drones. In most swarms, that's one leader and forty-nine followers. Which means one shot can take down all fifty. Episode 4: the difference between a real swarm and a light show, in under a minute. Zero jargon.
1
1
176
Over 450 missions flown, with pilots averaging more than 4 each. Must be something about TRY AGAIN. Congrats to PRIMAL MAVERICK for sitting on top of the leaderboard.
LAGRANGE (@lagrangedev) profile pictureLAGRANGE
@lagrangedevOct 22
3 seconds of thrust. Then you glide. Somewhere above you: a drone. Pursuit is a free browser game. Find it on radar, fly into it, get on the leaderboard. Post your T+ time. pursuit.lagrange.dev/?utm_so…
3
8
1,791
Load more