The whole of doomerism relies on whatever this is

This whole story is entirely overflown. It’s $1B over a 3 year period. BCBS plans have over $1 _Trillion_ a _year_ in medical costs. Isn’t even a rounding error.

Whatever your instinct says, try solving the problem with an agentic loop first. Stop thinking in code. Set up an eval or verifier, then run the loop.

Yeah this is the sermon of our times

Everyone should watch this, regardless of their domain. 心なき 身にもあはれは 知られけり 鴫立つ沢の 秋の夕暮れ

After Opus 5.5, I think it's time to retire "vibe coding." This is just how programming works now.

红杉的一位合伙人 Pat Grady 昨天直接把他们和自己的一个 LP meeting（Boston College）的内容录制发了出来，信息密度非常高，强烈推荐大家去看一下： - JEV 在过去的 7 天里从零增长到了 100 million 的收入。在此之前 The Information 刚刚爆出来，他们现在估值是在 10 billion。 - 现在 AI 公司的增速，比历史上的都要快很多。Instinct 发布以来，保持了 day over day 10% 的增长率。 - 200 到 700，指的是某一个 high-value knowledge work company 从去年年底 200M 增长到今年预测 700M 的收入。（后面那个 2 到 50 和 0 到 70%，他说还是跳过吧，可能还是有些敏感。） - 在 AI 公司，现在有一个连融两轮的行业惯例或风气：把"陪你建公司的"和"只出钱的"拆开。他拿红杉内部的数据（过去 12 个月里的 7 个案例）来举例： 红杉作为陪建的那一方进场，平均的投后估值是 1.1 亿；而仅仅一个月后，只出钱的那一方给的下一轮平均估值就达到了 34 亿。他自己的说法是以前从没见过，这就是泡沫。 - 很多企业想 own 自己的 intelligence：基础模型厂商只给你 Pareto frontier 上有限的几个点（大中小几档），但企业自己的 workload 需要的往往不是这几个点，所以要自己训、自己部署、专门为自己的 workload 优化。这也是为什么他说专用架构的 lab 在跑赢通用的 lab。 - 应用层公司每 4 个月就要 reinvent 一次自己，因为底层模型一直在变，地板一直在抬。 - 组织从 hierarchy、command and control 往 network of agents 走：AI 负责信息流转，人相对自主地工作。他说没有 Jack Dorsey 几个月前那篇帖子说得那么夸张，但方向就是这个。 - labs 内部：模型去年大部分时间已经在自己造自己，最近才意识到 alignment 没被当回事；都在押 custom silicon、新架构和 continual learning，算力全员紧缺；为了抢 API token，一边降价，一边派 consultant 去给 Fortune 500 定制吃 token 的工具 - 模型能力和实际落地之间有很大的 diffusion gap，这就是应用层的机会。 - hyperscaler 今年开始借钱做 capex，不再靠 free cash flow。 而且我特别喜欢他这种说话方式，他每句话就是只说一遍，说过去就过去，所以信息密度很高。推荐大家可以直接去看原视频， 总共视频时长只有 15 分钟：loom.com/share/c016702964a04…

That a billionaire would spend so much time doing something that any of us could do shows a remarkable equality of consumption in the modern world.

I’m so fucking sick of giving away 3% of all our revenue just for processing a payment. Like how is that not disrupted yet? Most ecom brands run on around 15% net. So we are giving up over 16% of our potential pre transaction fee profits to these bastards. Someone please disrupt this. It’s as if credit card processing companies are 16% equity holders in all our businesses.

RT @Brahmonaut: The best entry on Petrov is in encyclopedia Brittanica which says, “He is survived by his two children, two grandchildren a…

Thanks! We compared against a Momentum Prediction baseline, which is similar to JEPA. It was comparable in clean settings but collapsed in distractor and natural video settings. More work may be needed to make JEPA-like world models work well in these tasks. nitter.net/LunjunZhang/status/210…

Usually I can figure out how anthropic has done things from a mixture of my friends there and using the model but I have no idea how opus 5.5 is this good. Like, if I had to really guess it's just data, which is my Occam's razor for model improvements. But like, how

The thing is: from the beginning we always would have continued to call what we have now LLMs, and it’s not like you guys were saying “the LLMs you have now are stochastic parrots, but the obvious next step you guys are gonna do next is totally different”.

Thinking AI might be conscious is like thinking that there are real people talking to you from inside the TV, or that there is a tiny band playing music inside the radio. Fine for 2 year olds to believe, but beyond that, insane.

Adapt. Accelerate. Ascend. NEVER doom

"It's just a stochastic parrot bro, it just predicts the next token bro"

Andrew has been a consistently adroit voice in cutting through these arguments for many years now. Happy to say I’ve been an eager follower for a long time.

The phrase "stochastic parrot" is full of sound and fury, signifying nothing. Which, ironically, is exactly what the authors got wrong about language, and the core technical mistake of the paper. 1/

can someone explain how they got opus 5.5 to be this efficient & this good? i’ve barely run into usage limits at all. it’s absolutely cooking on genuinely difficult problems with what feels like zero effort while somehow also consuming dramatically less of my quota than other frontier models??! compared with astra or sol it’s night & day. the capability per token here feels kinda absurd, & this coming from anthropic where we didn’t see this barrage of usage limit resets or whatever.

Hang it in the Louvre