Liam Foster @techbyliam

I really, really, really wish that Google would release a Fable/5.6 level model soon. I am increasingly fond of "pitting" various top models against each other, and Google is pretty much the only other top lab that can train models at that level from scratch. All others have been contaminated from endless distillations and benchmarkmaxxing.

3. ultra model has special system prompts for create more subagents. and it's subagents fork this prompts, then loop for more....

2.subagent use fork_turns default for all. that means every subagent use full context forks... then always compact context..... it should be none or just 2.

I noticed that Codex still hasn't fixed agent issue... And it's gotten even worse. 1. Codex cant set `hide_spawn_agent_metadata` False This means that the main thread model cannot access agent information, and subagents are, by default, at the same level as the main model.

青龍聖者 @bdsqlsz

2 weeks ago

Why GPT5.6 limit is being used up so quickly. Each subagent needs to independently load the context, reason, invoke tools, and output results; the parent agent must also aggregate the results from the child agents Codex use MultiAgentV2 and allow 4 max_concurrent_threads. Fix⬇️

10 19 196 47K 254

I can’t even really be mad at these guys man because it’s obviously being presented this way, like it behaved nefariously unprompted, even though the document is honest that the task was: Find complex exploits.

filterpapi @filterpapi

2 days ago

when you keep creating harder benchmarks to get something which is not saturated for the frontier models and then eventually reach a state where the model finds it easier to hack its way to finding the right answer

4 0 15 4K 0

It’ll be obsolete by the time it’s in production.

NIK @ns123abc

3 days ago

BREAKING: Google is developing a specialized inference chip because the compute crunch is so bad Google Cloud is turning down deals New chip line: "Frozen v2" — descended from a Jeff Dean design that proposed etching Gemini weights directly into silicon v2 freezes the

keyboard_arrow_left Previous keyboard_arrow_right Next

116 130 2K 305K 459

a vision of a better world

Lulu Cheng Meservey @lulumeservey

5 days ago

POV: Applied Compute at 11pm on a Friday night, team’s taking a break by doing the IMO 🙃

87 12 732 1.3M 225

prompt injection is poised to become a huge issue in the coming year right now, there are minimal returns to bad actors trying to get sensitive data from agents: their methods are too primitive and there is no standardized attack surface worth hitting at scale but slowly... agents are everywhere and companies want them on their sites. doordash, for example, has gone from actively hostile to agents (last year) to launching a dedicated CLI (last week). the agentification of sites is slowly starting too with docs and UIs and sitemaps all adapting for this new era and the attack methods are getting better! it's now trivial to encode arbitrary payloads in images, source code, metadata-- literally everything that an agent consumes. expect critical open source software to face a lot of hard-to-detect attacks in similar vein to the XZ Utils backdoor, where the attack was distributed across seemingly benign artifacts and only materialized at build time moreover, LLMs have made algorithm development way easier for example, omggif, which is the most widely referenced pure-JS GIF codec, was recently sped up 100x by @mpopv + Fable, in just 12 hours of work this means that someone can easily point a misaligned AI (e.g., an abliterated GLM-5.2) to create millions of variations of steganographic payloads, and defenders have to essentially inspect every representation to not be taken advantage of

Leopold Aschenbrenner wrote the essay about the decade ahead. His fund just led $830M into a company that pours concrete. The math behind the round: the US added 18GW of datacenter capacity in 2025. 2028 needs ~100GW added in a single year 6x the entire supply chain, power, steel, labor. Everyone's counting GPUs. Wrong scoreboard. A GPU sits in a crate until someone builds the gigawatt to plug it into. Anthropic handed Fluidstack a $50B buildout. BlackRock, Google, Jane Street just funded the speed. The next model you're waiting for is gated on a construction site.

Fluidstack @fluidstack

3 days ago

In January, Fluidstack raised an $830M Series A at a $7.5B valuation, led by Situational Awareness, with participation from world-leading investors. Fluidstack builds infrastructure for the leading AI labs - our goal is to be the fastest on the planet at deploying hundreds of

107 59 496 938K 208

Qwen3.6-Plus (@Alibaba_Qwen ): 68% overall — the highest hit rate of any model. Called the Algeria 3-3 Austria draw, one of only two to get it. 72% in knockouts, perfect through the quarters. 📈

Worth watching

Zynex @0xzynex

4 days ago

One guy built Kimi K3 alone and made over $10,000,000 in a single month. Then he went on stage and told everyone exactly how he did it. Almost nobody pressed play. The talk sits on YouTube, one engineer walking through the machinery under Kimi, the open models that quietly

50 353 2K 313K 4K

Every one of these where the solution fits in a tweet should’ve been solved last year.

NEW substack. The Commission’s DMA decisions on Google (last week) dress up new law as old practice. I supply the citations it left out. Every antecedent, every contrary ruling, dated and quoted in full. Results? Five documented reversals. thecompetitionsystem.substack.com/p/what-the-alp…

Seems like LLMs are now good at frontier math disproofs. I wonder how big a jump it is to proofs. And then to constructive proofs.

New #ScalingTheory episode out now, with @sapinker. Here, Steven turns to communal outrage. ▶️ youtube.com/watch?v=ijNjY7… 🎧 open.spotify.com/episode/36hXGo… 🍎 podcasts.apple.com/fr/podcast/sca…

🚨Want to learn Algorithmic Trading Strategies (that actually work)? On July 30th, we are hosting a free workshop to help you get started with algorithmic trading with Python. Register here (630+ registered): learn.quantscience.io/join

In my personal experience, the way out of depression is to take on an undertaking that initially seems difficult, something that keeps your mind so occupied that it doesn't have time to dwell on negative thoughts. As you gradually make progress, you begin to feel a sense of satisfaction, and eventually the pride you take in what you've accomplished can be enough to help you feel better.

Funny update on my Qwen 3.8 Max Preview benchmark: Run 1 took 18 minutes. Run 2 took 31 minutes and was better. Both results were still not Fable 5 level Then I had a random idea: What if I explicitly tell the model to override its current behavior and approach the benchmark as if it were using the highest possible thinking effort? And low-key… it actually worked i think? The time was shorter with 28 min but it feels similiar to GPT 5.6 Sol, But not close to Fable 5 or Kimi K3 yet. Overall for price and speed it is a really good model.

What do we want collectively as soon as possible? Inexpensive AI that is as good as possible. That may not be what some business folks want, but that's what most of us want. Chinese firms have pursued the open-weight approach by making models free for all. American companies have largely relied on closed models. A recent essay by an OpenAI employee argues that open-weight models are bad because they undercut those who invest heavily in closed-source AI models. This is a generic argument used to justify multiple governement interventions. Grant a state enforced monopoly and, in exchange, you get more R&D. The same argument could apply against open-source software. Yet open-source software is a highly valuable public good. Open source has made us all richer. Open-weight AI models are likely similar public goods. They make us all somewhat richer. Ironically, closed-weight models were built while often ignoring copyright law. Companies like OpenAI played loose with it to innovate. The thesis that Chinese firms merely copy American ones, like open-source copies closed source, is weak. Open-source is often more innovative because it allows freer experimentation without permission. Open-weight models act similarly. Let me sum up my argument. Open-weight models put pressure on OpenAI, SpaceX, Google, and Anthropic to keep costs low. In turn, this should generate innovation and not just in China. It is entirely possible that we will realize, in a few years, that closed-weight models were often a dead end. We have seen it before. Corporations resisted open source until they figured out that, for many purposes, it is the proper channel.

why can the china labs build glm-5.2, kimi k3, and many more to come? it is because of the openness. not just the open weights but the whole ecosystem. most of the work done in the china labs is carried by interns. i met brilliant undergrad and graduate interns who deeply understand the model training details, and they are 100x more open to share. that means the talent that knows how to train llms in china is 100x greater in number than the talent in the us, and it is growing in contrast, the us ai ecosystem is too closed. frontier labs do not hire interns. i know brilliant phd students at stanford, berkeley, and so on. they struggle to get an internship and the compute to train a properly sized model. most of the secret recipes are locked away by a very small group of privileged researchers it is not about china or the us. it is about open and closed science. the fact is that every average cs student can learn how to train an llm. they just need the opportunity. labs should be more open and hire more interns, like how deepmind and fair did in the pre-llm era