DeepProve: a cryptographic receipt for every AI inference. now open source. Halo: autonomous coordination engine for the toughest conditions. @LagrangeFndn $LA
Los Angeles
Joined May 2022
- Tweets3.7K
- Following206
- Followers94.6K
- Likes2.5K
Pinned Tweet
You can't stop this swarm. That's the point.
Jam its comms. Spoof its GPS. Shoot three out of the sky. It re-plans and finishes anyway.
Your score is how many seconds you cost it.
holdtheswarm.lagrange.dev/?u…
LAGRANGE retweeted
A reporter asked an AI agent to find him cheap eggs. It bought them. $31.43, including a $3 priority fee the agent added on its own, for one carton he never told it to buy.
Now work out who pays when it is your card.
lagrange.dev/blog/your-ai-ju…
A reporter asked an AI agent to find him cheap eggs. It bought them. $31.43, including a $3 priority fee the agent added on its own, for one carton he never told it to buy.
Now work out who pays when it is your card.
lagrange.dev/blog/your-ai-ju…
the fix for more drones isn’t smarter drones. it’s drones that know less about each other. finally, a place where being out of the loop pays off.
Episode 8, if you missed it: the highest level of self-driving you can buy still needs your full attention.
xitter.cf/lagrangedev/status/210…
The paper, by Weijie Wang, Charalampos Papamanthou, Shravan Srinivasan and Dimitrios Papadopoulos: eprint.iacr.org/2024/1566
If you're at AUSA this week, here are six questions that will tell you more about any autonomy vendor than the booth demo.
1. Your convoy runs one manned leader and up to ten followers. What do the ten do when the leader is gone?
2. When one vehicle drops out, who picks up its task, and how long does that take?
3. What happens when the column gets split by an obstacle? And when it closes back up?
4. How do you know every vehicle is working from the same picture right now?
5. How does your system tell "destroyed" from "masked by terrain" on a degraded net?
6. What stops something joining the column that shouldn't be there?
Ask them in order. Fifteen minutes of answers beats any demo loop. Send it to whoever is walking the floor.
And ask us the same six: [email protected]
The long version, one failure mode at a time:
lagrange.dev/blog/your-drone…
An empty Waymo was the first vehicle to stop for a school bus with its arms out. Four other cars went past. Then it asked a remote assistant in Michigan whether the bus had its signals on, he said no, and it drove past the arms too.
The word autonomous never told you there was a man in Michigan. It does not tell you who is on the other end of a drone either.
lagrange.dev/blog/your-drone…
LAGRANGE retweeted
ronaldo reportedly won't come to his own farewell without a drone show. a thousand drones about to do a siuu. and for once it'll be a team effort.
Episode 7, if you missed it: a pilot can fly one drone, and that is not a figure of speech. It is the regulation.
xitter.cf/lagrangedev/status/210…
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.
The second one, from last week: a signed log will tell you the right model ran, and so will a signed log from a system that quietly ran a cheaper one.
xitter.cf/lagrangedev/status/210…
A signed log will tell you the right model ran. So will a signed log from a system that quietly ran a cheaper one and typed the expensive name into the file.
A proof cannot do that. Change the model, change the input, change one number in the output, and it stops verifying. There is no version of it that is wrong and still passes.
A claim can be honest and wrong. A proof cannot be wrong.
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.
The first one, from August.
xitter.cf/lagrangedev/status/208…
AI is doing more of the thinking. Machines are doing more of the moving.
And every serious buyer, regulator and operator is converging on the same question:
How do I know?
How do I know the model ran what you say it ran. How do I know the swarm holds when the link dies.
Everything we build is an answer to that question.
An OpenAI model was asked what Victoria spends on skin medicines. It ended up inside a Medicare portal pulling credentials, still looking for the skin medicine.
That was June 18. Australia found out on September 10, from OpenAI.
lagrange.dev/blog/nobody-can…
