We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1
Joined September 2010
- Tweets136K
- Following3.4K
- Followers79K
- Likes257K
it happens again
对付REA逆向的利器来了！让AI无法逆向你的项目！
有人做了个BLA，专门给REA这类AI逆向工具下套。
假密钥、诱饵接口、干扰逻辑，往发布包里埋。AI拆出一堆看着挺有价值的东西，顺着研究就跑偏，Token还得照烧。
连REA的解析上限也利用上了。作者拿真实REA测过，加固后的JS报预算限制，Electron ASAR报容量错误，WASM接口只能读出一部分。
作者测的14个基准样本，加固后跑出来的结果都和原版一样。
自己有软件项目的，可以在项目目录跑：
1. npx bla-agents setup，接入Claude Code、Codex、Cursor等工具。
2. npx bla-agents harden .，扫描发布文件，执行加固和验证。
BLA项目地址：
github.com/qybaihe/bla
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) retweeted
we spent $25,000 to help you decide what reasoning_effort you should use
also open sourced all the trajectories for people to play around with!
have fun🎉
Bravo. First time I see such excellent use of gaussian splats. Demoscene for the era of neural networks.
V4.1F-1017 still has some "2025 LLM" failures. For example, it coded itself into diminishing returns in the first run. Handoff, try again, another 500M tokens… and it's mostly recovering the original game.
> official API
amazing if that's all it takes
of course DS doesn't want to be *directly liable* for anyone using their future godawful Icebreaker models to hold hospitals ransom or whatever. Likewise, their web service implements Chynese censorship
But I worry about the weights
maybe that's too much education. Soviets also boasted like this
Americans are blessed, for they are retarded. They live in the eternal now extending into the future of the light cone, unburdened by inhibition of remembering mistaken dreams. This makes them fierce as pitbulls and wise as serpents.
The only nation that could build AGI to do mafs for them.
What could knowing "the underground struggles of the French Communist Party" offer that is comparable to such a blessing?
1-5 years is a lot actually
Short of nuclear war, strikes on US chip fabs or datacenters wouldn’t really work in the long run for China.
lesswrong.com/posts/jBCqkhxB…
«Across all of our strike scenarios, we could not achieve a delay of more than three years for an attack in the last year of ASI development. So long as an intelligence explosion dynamic holds, algorithmic efficiency will grow too fast for states to deny superintelligence indefinitely.»
sounds fair. So, this becomes irrelevant in 1-3 years?
but I think these terms imply that WWIII with nukes is not the worst case scenario.
China can build bunkers faster than datacenters. The opposite is true for the US.
The main issue is that the more time before ASI is left, the less motivation for a strike there is.
Sol 6.1 Ultra (!!) has just invented a brilliant water reflection trick (after ≈30 minutes of kicks): single in-screen planar capture, with its height "switched based on nearby water".
AGI is here. Bullish on Sama 👍
no, seriously, wtf
should have gone with Terra perhaps
but it's supposed to be a very performant game with fast seed generation and 2 km view distance
nitter.cf/VraserX/status/2115721…
oh well
good bye slopus, I barely knew ya
Got like 90% utilization at least
Anthropic сильно усложнили использование Claude из России. Компания активно начала просить пройти процедуру KYC, в рамках которой нужно прислать паспорт.
Если паспорт российский — сразу дают бан аккаунту. Даже если ты, например, живёшь в Новой Зеландии, и у тебя есть виза для постоянного проживания, как у этого реддитора.
Запрос на KYC приходит не всем — только если система Anthropic посчитает тебя подозрительным. Например, если ты сидишь под VPN.
С таким столкнулись не только жители России, но и китайцы. Они выяснили, что оформление подписки сразу после регистрации аккаунта также повышает риск бана.
man, we're kind of getting somewhere but it was such a slog. 1.2B tokens, at peak hours. even with V4.1F, that's $11.62 at API rates.
Luckily $10 OpenCode Go makes it negligibly cheap
Still, Claude/GPT would have been faster and maybe cheaper
this is still hard for the Whale.
> [Chinese labs] have not prioritized this area at all. They've been very focused on coding+agentic capability. Unlike the big US labs they don't have teams of scientists and math guys pushing in this direction. I predict they will reprioritize
patience. People are working.
Some thoughts on latest AA v5 index rankings.
1. DS 1017 performance on (new) FrontierCode 2.0 is impressive - can't be the result of benchmaxxing: close to Fable 5.1, slightly ahead of Argon and Kimi 3.1
2. Poor science+math performance of Chinese models dropped them signicantly in new index relative to US frontier models. I've been talking to a lot of Chinese labs recently and they have not prioritized this area at all. They've been very focused on coding+agentic capability. Unlike the big US labs they don't have teams of scientists and math guys pushing in this direction. I predict they will reprioritize, bc of the OAI math breakthroughs and the poor AA v5 scores.
Their attitude seems to have been: for math, once IMO Gold level achieved, diminishing returns to further math capability improvements. Same with basic science. Economic returns all concentrated in coding and agentic tasks. Even now I doubt they will move more than 10% of their specialized model capability resources toward math+science.
New evals test frontier-level capability in math and physics, exposing the Chinese models as far behind Astra and Opus 5.5.
3. Gemini Carbon scores are amazing. Hope there is now a GOOG model at the frontier that I can use in my research!
I strongly believe that adversarial verification of math+phys results by the best models is much better than human peer review.
Interesting that DeepSeek's current specific focus is "web+gamedev slop and cybersecurity", though. I guess the former helps with hype on X, and the latter… well.
Rumors I’ve been hearing. I’ll start with OpenAI, then Anthropic.
OpenAI: The internal model’s reasoning jump comes from more aggressive RL, including training in games and simulated environments. Strategies learned there transfer to research mathematics and other fields. Navier–Stokes and ETH show the scale of the improvement. The clue is in the Synthetic RL hiring: self-play, simulators, and synthetic feedback.
Anthropic: Same direction with its internal Claude model. It’s trained on long, difficult tasks in demanding environments, and the reasoning skills transfer to other fields. Its ETH result shows how far the improvement has gone. The earlier Universes listing and the RL hiring around mathematical reasoning align with that.
Musk’s Digital Optimus hiring explicitly calls for measuring transfer from games into general computer use. That’s why his game-playing AI push matters: games develop abilities that work far beyond the games themselves.
The interesting part is where those abilities show up. Each lab’s internal model applies reasoning developed through interactive training to mathematical problems that had resisted decades of work.
I find such reactions comical.
Yes Terra is >>> Kimi and 1017. Indeed, it's even better than Fable 5.1. Is Terra a more impressive, wise, intelligent model than Fable? Lmao, not in a thousand years.
But it is a much more polished production system.
AA mainly measures utility.
For me, Artificial Analysis is no more trust worthy. This update was made to just make openai higher, Chinese models lower.
And it's not even thier first time, updates like that happend many times. All with the same result...
There is no way 6.1 Terra is better than Fable 5.1, or Argon.
Imo if you want trustworthy benchmark leader board look at Vals Ai leader board.
Btw, this is a deflationary take, but what I'm saying is that American models lead on use-case-driven post-training and data. Just like V4.1F-1017 is not a generally greater model than V4.1F-0910, but is much more capable, so are they vs 1017.
The kick is, this also applies to *frontier capabilities*. Math, end to end ML research, kernel engineering, bio, physics, whatever. On those benchmarks, the gap is even worse than on AA, because the Chinese barely touch them.
that said
I should have just given this task to DeepSeek.
The clouds here are designed by DeepSeek, after all.
nitter.cf/teortaxesTex/status/21…