Anthony Leverrier

Started to curate a table of small quantum Tanner codes aleverrier.github.io/qtanner-search…

well-deserved Turing award for Bennett and Brassard!

So I think it's becoming increasingly clear that efficiency and losslessness, across both compression and decompression, together represent four potential axes along which we can begin to parameterize the space of possible (intelligent) minds. But what are the others? (12/12)

I think one of the conclusions we should draw from the tremendous success of LLMs is how much of human knowledge and society exists at very low levels of Kolmogorov complexity. We are entering an era where the minimal representation of a human cultural artifact... (1/12)

Based on the data I see, I think: - Anthropic🇺🇸/Google🇺🇸/OpenAI🇺🇸 all ~tied - Meta🇺🇸 / xAI🇺🇸 each ~7mo behind - Moonshot🇨🇳/- Deepseek🇨🇳 / zAI 🇨🇳 / Alibaba🇨🇳each ~9mo behind - Mistral🇫🇷 ~1.5 years behind - No other companies competitive

I keep reading this take (below) every few months, presented as if extremely profound, and it is just offensively dumb. It confuses data and information, it ignores the fact that not all information is equally valuable, and it ignores the importance of retention rate. As a thought experiment: if this were true, if your retina cell count were 10x greater, you'd be "trained on 10x more tokens" and therefore you'd be way smarter. Same if their firing frequency were 10x greater. With 10x more retina cells firing 10x faster you'd be "trained on 100x more tokens"! Obviously this makes no sense -- the signal coming from these cells is extremely correlated over space and time, so their raw information content (what remains post-compression) is extremely low compared to the "raw bit" encoding. The human visual system actually processes 40 to 50 bits per second after spatial compression. Much, much less if you add temporal compression over a long time horizon. Latest LLMs get access to approximately 3 to 4 orders of magnitude of information more than a human by age 20 (post compression in both cases). About O(10T) bits vs O(10-100B) bits. And that's just *raw information* but of course not all information is equal, otherwise we wouldn't be spending tens of billions of dollars on training data annotation and generation. Plus, that's only *information intake* but of course humans have far lower retention than LLMs (by 3-4 OOM). You could write a short essay about how incredibly off the mark this take is.

We just went from horse-and-carriage to car. The old manuals still have truth in them, but the whole map of effort, speed, and leverage changed overnight. Building used to mean holding an entire system in your head. A fragile memory palace. A house of cards in biological primate RAM. If you stop, the palace collapses. Dinner, a meeting, a context switch. You come back and it’s glass dust on the floor. That’s why builders can look “antisocial.” It’s not vibes. It’s survival. You’re trying to get the palace out of your head and into code before it evaporates. Then the weird miracle: I type a few paragraphs that barely make sense on reread, and the machine builds the palace anyway. It mirrors the structure. It fills the gaps. It hands it back. The feeling is not “wow productivity.” The feeling is: I am seen. Like the part of you that has been translating yourself for 20 years finally gets understood on first contact. This is why the new skill is not “code faster.” It’s taste, direction, and leadership. Managing a swarm of agents. Running tight loops. Knowing what to ask for. And it brings back something old-school: apprenticeship. We forgot how to teach. Now teaching matters again, because the tools are insane but the mind behind them still has to be trained.

Some people seem to think generating 1000 papers in a couple of months is a lot. The question, imo, is whether an ecosystem that generates and evaluates one million papers every second is desirable (has positive ROI for society) Machine-speed is different

True but you’re missing all the fun in the meantime

We estimate that Claude Opus 4.6 has a 50%-time-horizon of around 14.5 hours (95% CI of 6 hrs to 98 hrs) on software tasks. While this is the highest point estimate we’ve reported, this measurement is extremely noisy because our current task suite is nearly saturated.

@beffjezos just a sigmoid!

Can you feel the acceleration?

Every word of this, yes. By @tylercowen

@75Jamin @alex_avoigt 75% of French electricity is nuclear

@alex_avoigt Why this big difference in 1990?

Solar and Wind win against Coal.

Pretty happy to see this paper published, and I just realized it was the 2000th paper in Quantum quantum-journal.org/papers/q-2026-…

@SingularIV66301 @olivez @IonQ_Inc Much slower hardware though

@olivez @letonyo @IonQ_Inc already reached 99.99% fidelity last year. 99.9% isn't a thing now.

amazing new benchmark for quantum computing. Now we just need to figure out how to decode qLDPC codes very well! arxiv.org/abs/2602.11457

Very excited not just about this challenge, but what it signifies in terms of models ability to generate new science. This is only the beginning!

@DotDotJames I'm searching for these codes quite systematically, pretty much as explained here: arxiv.org/abs/2512.20532

@letonyo very curious. are you discovering & testing these codes, i.e., first find, lit search / complement to ECC zoo, etc? partial context; hypertokens exploit Tanner codes and esp. those which are CRT-based and so of general interest, been considering quantum sims using same, etc.

Started to curate a table of small quantum Tanner codes aleverrier.github.io/qtanner-search…

@olivez that's the second strong assumption of the paper (besides the existence of fast decoders)

@letonyo and to build 100K qubits with 99.9% fidelities :) !

timelines keep getting shorter

The Simons Foundation has put their treasure trove of videos on YouTube (they’ve been available for a decade from the Simons website, but I kept begging them to move it to a platform with many more eyeballs…). Here’s Eli Stein interviewed by Charlie Fefferman: youtube.com/watch?v=pWOU-c… And Rutgers’s @RutgersU own Endre Szemeredi (interviewed by Avi Wigderson): youtube.com/watch?v=_ZQ4aB… And Ron Graham (here talking about Math+AI, from a decade ago): youtube.com/watch?v=fHrply… And lots of Gromov, e.g.: youtube.com/watch?v=aI2ASq… Sorry for wasting the next few hours of your life, but it will have been well worth it to watch these!!