#大佬观点

Thoughts About Scaling Law Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions. The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — roughly 2.7:1 — and the industry complied: GPT-3, Gopher, MT-NLG. Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter, and that with sufficient compute the two should grow at the same rate rather than drifting apart. The error in the earlier fit compounded with every order of magnitude of compute, which is why the largest models of that generation were the most misallocated. The trillion-parameter round was, in retrospect, a detour the whole field took together and then reversed. Chinchilla wasn't the end either. It optimized training compute for models that would be trained once and evaluated. Today a model is called billions of times a day and inference dominates lifetime cost. Put inference into the objective and the optimum moves toward smaller models trained far longer — deliberate over-training, which is what Llama-2-7B and Gemma-2-9B were doing at roughly 290 and 889 tokens per parameter. Sparsity moved the target again. In a MoE model two quantities have to be kept apart: total parameters govern roughly how much the model can hold — knowledge, facts, the long tail — while activated parameters and effective depth govern roughly how far it can think, how many steps of a causal chain it can carry before it comes apart. A dense 20:1 ratio does not transfer. And the ratio isn't a single number at all: Roberts et al. (2025) find the optimal tokens-per-parameter is task-dependent, with memorization favoring more parameters and reasoning favoring more data. Follow-up work on MoE observes that at fixed TPP, pushing total parameters higher actually degrades reasoning, while activating more experts reliably helps it. This matters for what we are building toward. Finding a vulnerability is not a retrieval problem. It doesn't come from having memorized more CVEs; it comes from carrying a twenty-step chain of inference to the end without losing the thread. That capability does not live in total parameter count. Which brings us to this release. Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal. Well, scaling has more than one dial. We turned the post-training one this time because it had the most slack left in it — not because the others are finished. Base model size, pretraining data, compute spent per forward pass: all of them are still on the table, and we will come back to each. What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling. Next time, maybe mid-training, pre-training, and even more.

Travis Kalanick: "A lot of folks think I'm back. I've been working my ass off the whole time. I just haven't been talking about it." Eight years, thousands of employees, multiple industries, none of it on LinkedIn, and now the biggest check Ben Horowitz has ever written. @travisk joins @bhorowitz and @eriktorenberg for a fireside chat to discuss Atoms: an industrial AI company that sees manufacturing, real estate, and logistics as the CPU, storage, and network of the physical world: 00:00 Intro 01:30 The Uber video from 10 years ago 04:26 Bits, atoms, and the three computing primitives 07:37 Can a delivered meal cost less than the grocery store? 10:16 Standing next to a TCP packet 11:59 Ride-sharing was the gold medal. Transport is full of silver 13:59 Resistance to change is the final boss 22:22 Why he didn't buy Lyft 24:09 Best idea wins 28:13 Blood, sweat, and ramen 32:32 When it gets easy, push harder 36:18 Showing up to the fundraise like a guy selling watches 45:46 Dirty fuel vs. falling in love again 48:31 Be uniconic

Qwen 27B is the "DeepSeek moment" for open source. It matches the closed-source state-of-the-art from just a few months ago and runs on an RTX 5090. Without exaggeration, it’s a game changer.

Physics is going to be as cooked/cooking as math. I fed Claude an open problem in stochastic thermodynamics of the kind I'd suggest to a mathematically inclined grad student. And over a few days of back and forth, it did months of work and closed the whole problem class.

@trq212 because the creative aspects we care about are at the level of intelligence rather than technique

let that sink in

defenders can see the future, and have a narrow window to uplevel their cybersecurity practices now. key is to uplevel fundamentals and apply the best AI tools. what we’re doing at OpenAI, and where other organizations can start: https://blog.gregbrockman.com/the-defenders-window

When language models first started using tools well, I was sympathetic to the narrative that instead of scaling up language models, all we needed was a strong enough "cognitive core", say 1B parameters, and anything else could be done with tool use, like browsing the internet or executing code. I think a lot of people were sympathetic to this argument, and indeed it is pretty hard to come up with a meaningful task that cannot be in principle achieved by a 1B model with adequate access to tools. For example, any esoteric fact that a large language model would know can be, in principle, retrieved from the internet and reasoned over by a 1B language model. However I now think this is totally wrong for one simple reason: doing tasks quickly and naturally without tool use matters a lot. The way that I internalized this reason was actually in my personal journey learning badminton this year. In badminton I am very much like a "1B cognitive core". While I can physically do every movement in a badminton shot that my coach teaches me, it requires a lot of work to mentally remember every cue and put it together. In practice I can do a shot almost perfectly, but I struggle to do it across a point and I definitely can't do it consistently in a game. This is obviously different from someone who has practiced a shot ten-thousand times and effortlessly executes it as a natural instinct. In the same way, language models knowing a fact internally, without tool calls, is meaningful. The first reason is that we obviously care about speed; you'd much rather get an answer immediately than have the model think a long time to be sure of its answer or browse the web. A second reason is that there are some things that are simply best learned via backpropagation over lots of data. If you ask about how people generally think of the Shambhala music festival, you'd rather a large language model give you an aggregate opinion based on all the data on the internet, than get a regurgitation of the first three reviews that show up in a web search. A third reason is that having to do a lot of work to find an answer is not as reliable as already knowing the answer. While this does not have to be true in theory, it is probably true in practice, at least for now. If you have to re-look up facts or redo a mathematical derivation all the time there is a higher chance of mistakes, which can compound in a long-horizon task. Once you buy that it is valuable to do things parametrically without tool use, then you must buy the argument that a 1B cognitive core is not sufficient. There is an information limit to how much knowledge can be internalized by a 1B model, and we will surely want AI to know more than that. Even 1T probably won't be enough. We will want the AI to know as much about our world as possible, we will want it to be updated with new information, and our expectations of what AI can do for us will continue to grow. In summary, tool use enables small models to do a lot more, but those who demand the highest quality intelligence will always want larger models. Bitter lesson strikes again.

Teaching Everyone to Fish for Tokens Nvidia wants you building your own model, not buying from Anthropic/OpenAI. https://www.interconnects.ai/p/teaching-everyone-to-fish-for-tokens

Some thoughts on Dario’s post: 1. Dario does not actually address Gavin Baker’s account of what he said – something he could easily deny if it were inaccurate. 2. Dario claims his critics live in a “bubble” where all regulation equals regulatory capture. He calls this an overly simplified view and notes that “Many people outside this bubble think of regulation as something that constrains corporate power and benefits ordinary people.” This argument is a straw man. Of course treating all regulation as capture would be overly simplified – but almost no one holds that view. I have repeatedly argued for strong antitrust enforcement to keep industries competitive, especially Big Tech. If Anthropic continues toward monopoly or duopoly status, I would be among the first to demand those rules apply. 3. Regulatory capture is not vague or in the eye of the beholder. Nobel laureate George Stigler defined it as regulation acquired by an industry and designed and operated primarily for its benefit. Stigler challenged the traditional view that government regulation arises from a benevolent state protecting the public from market failures. Rather, industry groups have concentrated stakes and pour resources into influencing regulators, whereas the public’s stake is diffuse and unorganized. The revolving door between companies and the agencies that regulate them compounds the problem. Anthropic understands these dynamics: it has hired multiple senior Biden AI-policy officials and built a substantial government-affairs operation plus a network of aligned organizations to push its preferred frameworks at state and federal levels. 4. Dario has consistently pushed for a new federal agency to review and approve frontier models prior to release – a proposal framed variously as an “FDA for AI,” an “FAA for AI,” and most recently a “FINRA for AI.” I call it a “DMV for AI” because a review process modeled on the FAA or FDA (which takes years) or FINRA (which issues rules for a staid industry widely seen as protecting incumbents) will create long queues as AI models wait for testing and approval. This process will only become more labyrinthine as rules accumulate to prevent theoretical harms. This would handicap the U.S. relative to China, which will not adopt the same constraints. It would also undermine Anthropic’s own business model, whose pricing power depends on remaining ahead of open models. Whatever Dario states today, it is difficult to believe the company would simply accept outcomes that erase that advantage. 5. Anthropic is on track to become one of the most valuable companies in history, with the resources to navigate any approval process and shape the rules while competitors wait. Dario wants open models under heavier scrutiny – he has called them dangerous in Senate testimony, criticized them for not being centrally monitored or withdrawn, and linked them to IP theft. He says he has never sought a ban, but he could achieve a similar result by insisting that identical rules apply to both open and closed models. The U.S. risks becoming an island of costly closed models while the rest of the world races ahead with broader choice. 6. Dario acknowledges that AI is structurally centralizing but attributes this mainly to chips and scaling laws. Access to compute matters, but the deeper risk is who decides which capabilities are available to whom. His preferred pre-deployment testing and FAA/FINRA-style oversight would place that gatekeeping power in a federal bureaucracy working hand-in-glove with a small number of frontier labs – reinforcing centralization rather than countering it. 7. The second part of Dario’s post assumes we have amnesia about Anthropic’s well-orchestrated campaigns hyping AI fears. His May 2025 claim that AI would wipe out 50 percent of entry-level knowledge jobs within five years still lacks supporting evidence fifteen months later. Similarly Anthropic breathlessly promoted its heavily contrived “blackmail” study on 60 Minutes. Yet Dario blames public negativity on a long-standing loss of trust in institutions rather than his own messaging. 8. These narratives have done more than anything to shape public fear. People are left asking the same question Mark Zuckerberg posed: why race to build a future you describe in such negative terms? Thomas Sowell’s "The Vision of the Anointed" captures the mindset – elite intellectuals convinced that only they are enlightened enough to control the outcome. As Zuckerberg notes, concentrating power in the hands of an enlightened few has rarely produced the promised results; the practitioners turn out to be less enlightened in practice than in self-conception. 9. Gavin Baker summarized the disagreement cleanly on our pod: Dario believes frontier AI is too powerful to distribute; we believe it is too powerful to centralize. Dario appears to believe, sincerely, that safety and progress are best served by centralizing authority in a marriage of corporate and state power. The weight of human history gives us reason to fear that outcome.

locked in for the next 6 weeks