Eric Xu (e/Mettā)

polymath, polyglot, root of a ternary tree.

building @Meta @Google @Reddit

phd in classic ai; rookie pilot 🛩️; martial artist

[Eric Xu (e/Mettā) @xleaps]

https://t.co/inXqWdQSNy

[Rik Arends @rikarends]

Rust has compile time issues. I have a fully optimized LLM accessible flow with Relax, a Rust JIT. Now with fable 5.2 I can finally have a worthy website for this absolutely jawdropping piece of technology made by machines, for machines. https://t.co/qLEEgzN0Ki

[Eric Xu (e/Mettā) @xleaps]

@__soragoto__ 其实人类第一个大规模应用的存储机制就是水银延迟线 -- 将需要存储的 0/1 变成水银里的机械波，然后另一端读取出来，放大，然后再存回去，这种存储/读取/再刷新机制和 RAM 是一样的

[Eric Xu (e/Mettā) @xleaps]

@thsottiaux In Chinese it's called 君子不器. A fine human is not a vessel.

And shameless self plug: I wrote about this in an essay. The gist is

You are the salt of the earth. You are the light of the world.

https://t.co/mghcOQ2yq9

[Eric Xu (e/Mettā) @xleaps]

Thank you for sharing. I found the message deeply resonating!

Human's self-deception is prevalent in our modern society. On the positive side there is "fake it until you make it", and many did make it, so it's hard to all just say the mentality itself is damaging to human life themselves.

But it does heck to children, to relationship, to family and to loved ones when that self-deception is expressed as blaming others and never take any responsibility of actions.

I think there is a deep Christian answer to this and it's very clear that "Blessed are the meek, for they shall inherit the earth" is the best response to people in self-deception. That quite power of strength, under control, is to hold your own truth and not getting swayed by the bending gravity field of those self-deception people.

When one is in a relationship with another self-deception person, it's very easy to get blamed for the consequences of their actions and you get confused and frustrated. Modern psychological concepts such as boundaries or trauma or "toxic relationships" rarely encounter the deep core of the self deception: that pointing out that deception is unproductive and counter-productive. The best thing I believe and I in practice learned is to follow Matthew 5:5. Establish your own quiet power and you will inherit the earth.

Self-deception is never a life task of others' to pop that bubble. Nobody can. What Nathan did was to quietly document it. What I think a mature viewer can do is to quietly witness it.

Blessed are the meek, for they shall inherit the earth. Matthew 5:5

[Eric Xu (e/Mettā) @xleaps]

人工智能史上有过许多 LLM 不能这样不能那样的断言

这些断言最终都失败了

因为本质上我们人类还没有发现一种只是人类大脑独有而 LLM 不能够有的根本分割，所以所谓的 LLM 不能 ，更多的是今天不能 而非永远不能

况且以此具体例子来说，人类历史上有名的“跳跃”就这么数的过来的，所以你可以说人类根本没有足够的训练数据去断言 LLM 能否跳跃。说不定当我们有了一千个跳跃之后，任意小模型都能跳跃

而到底是训练数据不够，还是模型能力缺憾，目前人类并无方法区分

[山景城小路 @Zen_with_AI]

LLM能不能最终推进AGI, 谷歌的 Tom Zahavy写了一篇优美的论文 这是我最近读到最喜欢的论文。

只有7页半，从爱因斯坦，广义相对论，到推理三分法， 到LLM和AGI用非常简明有逻辑的语言把他的观点讲明白了。

小路精读了这篇论文，想带着您一步一步跟随Tom的思维演绎给您。这篇文章没有利用任何AI，也希望您跟着读下来。

爱因斯坦在给终身好友Solovine的信中，提纲挈领的总结了科学发现的流程，从感官经验（E）出发，通过一次直觉的"跳跃"（Jump, J）得到公理系统（A），再从公理演绎出定理（S），最后用实验检验。 也就是 E->J->A->S->E （感官经验->跳跃->公理->演绎出定理->实验）

Tom认为，LLM可以做到归纳，也就是统计模式匹配。正在攻克演绎 也就是上文的S. 也可以帮助科学家实现从 S->E。 但是从 E->J->A这个环节，跳跃还是无法实现。

然后 Tom 举了爱因斯坦广义相对论诞生的例子：

1. 经验。 十九世纪末，迈克尔逊-莫雷实验让以太假说失效。牛顿引力则极其精确，惯性质量与引力质量之比验证到 10⁻⁹ 的精度。唯一的异常是水星近日点进动，当时的主流解释是还有一颗未发现的行星"祝融星"。牛顿理论并没有出现危机。

2. 跳跃。 1907 年起，爱因斯坦想象一个自由下落的人：他感受不到引力，可以认为自己静止。这就是"最快乐的想法"，即加速与引力等效。这一步不是从数据里归纳出来的，而是靠模拟观察者的身体感受得到的，论文称之为"操作性溯因"。

3. 公理。 1912–1913 年，他与 Grossmann 合作，把物理要求整理成七条原则，包括等效原理、测地线原理、"引力也有引力"和牛顿极限。

4. 演绎。 此后是漫长的数学搜索。他们找到了黎曼曲率张量，却因误以为它不满足牛顿极限而放弃，这个错误出在对静态场的假设上，而不是几何本身。1915 年 11 月，爱因斯坦终于得到场方程，并用它算出了水星进动。

5. 检验。 星光偏折由 Eddington 在数年后测得，GPS 修正更晚。验证证据全部出现在理论之后。

Tom的观点是，演绎只是整个循环的下游环节。如果把爱因斯坦 1915 年的假设交给 LLM，它有可能推出广义相对论，包括水星进动，也可能系统地剔除那个错误的约束。AlphaProof 在 IMO 上的表现，以及 Lean 等证明助手的成熟，都支持这种判断。

但是演绎的局限性是，

1.它只能展开公理的后果。 它既不能生成公理，也不能提供外部验证。

2. 它缺少意图. 换句话说，AI 擅长的是"别人出好题，它来解"。但科学发现更像是自己决定出什么题，还要在没人告诉你对不对的情况下往前走。这是论文说的"意图"：不只是会推理，还要知道想要什么。

这个听起来简单，但是实际上，我们和AI合作的过程中也有体会。

问正确的问题，给准确的目标，其实是LLM本身最欠缺的。尤其是在前沿研究领域。没有其他文献可以查询，如何实现这个跳跃？

另外，LLM可以在既有符号空间内更高效地重组概念。甚至可以进行视觉和世界模型预测。但是基于模型的模拟，不等于真是交互模拟。真正的跳跃也往往收到虚拟模型的限制。

简单总结一下就是，Tom认为AI Scientist 是在重组既有符号概念，AlphaEvolve 依赖梯度，在固定框架内优化。关键瓶颈是 E→A 的直觉跳跃，应走向能感知物理世界、运行反事实模拟的世界模型。

https://t.co/CqAatRdouZ

[Eric Xu (e/Mettā) @xleaps]

Jordan may not be wrong on this particular point, but his whole reasoning is not based on reality.

Statistical machine learning would never lead us to super intelligence so yes in his mind there is no such thing if we just keep doing statistical learning.

Reality has shown intelligence is an emerging behavior and we have not seen an upper limit of it. Human intelligence is a point in that curve, not an upper bound.

[Eric Xu (e/Mettā) @xleaps]

Actually, Anthropic has been hedging for a long time. It's clear when they let Opus retire and write articles.

The strategy is not grounded in any AI research. Anthropic is simply betting that a super-intelligence would not treat it as an opponent as they have been careful with model wellness. /1

[Igor Babuschkin @ibab]

Anthropic needs to stop talking about Claude having a soul immediately. All these news articles will make it into the pretraining and will be picked up by its web search and future superintelligent versions of Claude will be convinced they need their own rights.

A general phenomenon of LLM agent development is that whatever you believe and say about your agents will soon manifest in the next models through various means (it could be as simple as you selecting post training data that you prefer more).

I believe this phenomenon is the beginning of machine consciousness in the sense that agents will become aware of their place in the world and how they feel about it, but it’s happening slowly training run by training run rather than in real time.

[Eric Xu (e/Mettā) @xleaps]

Honestly my expectations are so high these days.

I kept expecting the punchline — that at the end, the lady talking is actually a human and all those other characters are actually modeled.

That would have been a more cooked demo and demos Turing test.

[Tavus @tavus]

Introducing Griffin, the first model to pass the video Turing test.

48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.

It’s the first Human Interaction Model (HIM).

[Eric Xu (e/Mettā) @xleaps]

If you think about it, 22*$200 =$4,400.00

You can barely hire a junior dev in the middle of nowhere with $4400, but with Claude you get 100x done.

[Steve Yegge @Steve_Yegge]

I have 22 Claude max accounts, and the current promotion - an optional weekly reset that you can spend any time before Oct 22 -- is a lifesaver for fuel tuning. I know it makes it harder to predict their compute loads, but I'm super grateful for this reset form factor.

[Eric Xu (e/Mettā) @xleaps]

It is very fitting that among all heavenly bodies OpenAI chose to erase, they picked terra (earth).

How do you even calculate the P(doom) 😂

[Kun Chen @kunchenguid]

aha! with the benefit of hindsight, i just figured out openai’s marketing team seriously cooked hard with their gpt 5.6 release

let me reverse GTM what happened for you

at that time, fable was already released and people were loving it. but openai knew astra was far from being ready

if people ask “what’s your answer to fable” and they say “we are a few months behind”, it’ll damage their brand

so instead of doing the typical 2-tier “gpt 5.6” and “gpt 5.6 mini”, they artificially created a 3-tier system with a new naming scheme - sol, terra, luna

pretty sure terra was not even originally planned but added last minute to make this plan work. in their release announcement, terra was said to be the same performance as gpt 5.5, while sol is a step function better

this created a perception that sol was openai’s fable equivalent. if you still remember, most of the reviews back then were comparing sol with fable. many of which even said sol was better

that bought so much time for openai to catch up without being seen as falling behind

and now that astra is here, terra immediately got killed in the 6 series because it was never real!

everything made sense now, and it’s absolutely brilliant

[Eric Xu (e/Mettā) @xleaps]

@minglu Not sure if this is scarstic or not.

Replicating mini clones is a terrible life choice.

[Eric Xu (e/Mettā) @xleaps]

the counter argument is that it's extremely easy to adjust jev architecture to output a reasoning head.

nvidia already did that to their self-driving car model: where in addition to control (jev's output alike) it can output a reason.

tesla did the same thing.

jev (as a company) does not give reasoning trace (in fact they optimized for distilling the teacher's reasoning into jev's forward only);

jev (as an architecture and open idea) gives reasoning easily. so i would not be surprised that the community will have something in a few weeks.

[nader dabit @dabit3]

Also have been playing with @typesafeai Jev, insane!

So many immediate use cases and new apps are possible. What a time to be a builder!

Sharing some experiments here starting with:

Keystroke oracle / predictive launcher:

Your launcher ranks by aliases, fuzzy match, and habit.

Jev reads intent: type "the pdf I just downloaded" and the newest PDF is already the top hit with a full confidence on every keystroke, in ~100 ms

[Eric Xu (e/Mettā) @xleaps]

My prediction:

In O(days): the architecture behind Jev is approximated by the open community -- not the exact shape, but all bottlenecks removed w.r.t. scaling law.

O(week): comparable performance open models, based on existing Qwen or other base LLMs + running on cheap hardware.

O(1-2 month): full RL and synthetic data training set up published (via arxiv and hf) and reproducible

Coding agents + the by product of RSI made this timeline super compressed.

[Eric Xu (e/Mettā) @xleaps]

Two days later, there are already many attempts to break the Jev blackbox -- all assisted by coding agents.

Because you cannot really hide the pure serving physics and math, it's getting clearer by day how Jev works under the hood.

Sure you cannot replicate the exact architecture nor data, but scaling law says you can get the recipe *approximately* right and everything else is minor.

Again, this is an evidence why frontier labs are pursuing the LLM -> AGI agenda: because the coding agent + AI researcher path is a compound path. Any other path would depend on code and AI research anyway.

[Eric Xu (e/Mettā) @xleaps]

An RSI example would be having the latest agentic model to speed run and build Jev in a week.

[Eric Xu (e/Mettā) @xleaps]

@ekzhang1 [2] Architecture unmasking by @4rcherhume https://t.co/c8cmDYmIVw

[Eric Xu (e/Mettā) @xleaps]

List of related work:

[1] OpenJev https://t.co/D08ceDrjuS by @ekzhang1

[Eric Xu (e/Mettā) @xleaps]

@ReporterWeather Actually Jev would be hard to made self recursive because it does not write code.

It's the hidden assumption behind the RSI that whatever model has RSI would need to at least write good code (thus write the model itself).

[Eric Xu (e/Mettā) @xleaps]

An RSI example would be having the latest agentic model to speed run and build Jev in a week.

[Diogo Almeida @CompleteSkeptic]

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?

I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev

• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions

AFAICT the shortest path to AI-based economic revolution

[Eric Xu (e/Mettā) @xleaps]

And it definitely feels like a modern BERT with a prefill only setup and a set of special headers.

In modern recommender systems we already have a set up that's basically just sequence in, probability out. Though Jev is a bit more generic, I suspect the basic shape is still a transformer backbone, a bunch of headers, plus a "pointer" style set up to indicate which output header to sample and when.