triple_minded

Bring up the Right Questions | Not Investment Advice

[triple_minded @triple3minded]

Model routers weaken lock-in — Ramp cut AI spend 40% by switching models per task, raising the question of "intelligence saturation".

[triple_minded @triple3minded]

The AI capex financing number to file: Goldman now expects $420bn of hyperscaler gross debt issuance in 2027, up more than 60% from their $250bn 2026 estimate — about half of new Treasury long-duration borrowing next year. AI capex is becoming a bond market story.

[triple_minded @triple3minded]

European sovereign spreads are cracking. French 10-year up 11bp to 4.56%, the highest since 2008, and the French-German spread widened to 100bp for the first time since 2012.

[triple_minded @triple3minded]

Now the part that makes the second half of his quote follow: long-horizon tasks are only trainable this way. If an agent runs six hours over 200 tool calls, a human grader also needs six hours. Human signal doesn't just get expensive there — it becomes arithmetically impossible. So "讓模型自己記憶、自己進化、自己完成長程任務" isn't ambition, it's the only affordable route to that capability class. Chat was the last regime where human feedback could keep up.

[triple_minded @triple3minded]

唐傑為清華大學計算機系教授，他在清華大學的名為「機器學習」的課堂座無虛席，連教室走廊都站滿聽課的學生。16周的課程內容為大模型全鏈路訓練「一條綫」，「從基座、Scaling、合成數據、偏好與RL，再到Agent工種、持續學習、AI訓AI」。

他在課中表示，人類每退出一個閉環，訊號就更便宜一次，即每當人類不再直接參與某個訓練環節，改由AI自動完成，獲取訓練訊號的成本就進一步降低。他續指，AGI（通用人工智能，Artificial General Intelligence）及ASI（超級人工智能，Artificial Superintelligence）將是AI模型接下來的主要目標，「Chat這一仗基本打完了，2026年該做的是讓模型自己記憶、自己進化、自己完成長程任務。」

[triple_minded @triple3minded]

The logic: a model only improves if something tells it which outputs are better. That "something" used to be a human at every stage, priced in labour. Each time a loop closes, a per-item human cost is swapped for a per-item compute cost. Human annotation gets more expensive over time (wages, and expert scarcity); compute gets cheaper. So every closure is a permanent one-way ratchet.

[Z.ai @Zai_org]

We’re sharing how GLM-5.3 helped build and optimize the inference infrastructure serving GLM-5.3-Flash.

The system went from its first successful run to production readiness in less than two weeks, with end-to-end throughput tripling relative to the initial baseline.

The key was dense feedback: local correctness tests, execution traces, microbenchmarks, and end-to-end measurements that enabled targeted hypothesis testing rather than reliance on aggregate performance metrics alone.

https://t.co/yUf6OpJD7c

[triple_minded @triple3minded]

DeepSeek 研究员的判断：AI 只会把社会推向两端——要么生产力被极大解放，接近共产主义；要么是《赛博朋克2077》，少数几家公司垄断最强的 AI，没有它你连往上爬的梯子都摸不到。他不相信 OpenAI 和 Anthropic 会主动把能力分享出来，所以他选择留在开源这一边。

https://t.co/fbFZmLwKyR

[triple_minded @triple3minded]

Webster's strategic point is the episode's sharpest idea: technology confers no lasting advantage since everyone gets the same tools, so it amplifies or undermines existing edge. Invoking Christensen's law of conservation of attractive profits — if IQ is under attack, value accrues to EQ (trust, relationships, patience).

https://t.co/Ew2JRQQ57A

[triple_minded @triple3minded]

Scale figures Brockman cites: ~1.1 billion weekly active ChatGPT users, ~100 million in the US (roughly a third of the population), and 300 million health queries per week. Strikingly, he estimates another ~1.5 billion people have tried ChatGPT and stopped using it.

https://t.co/E6vb8ygTwx

[triple_minded @triple3minded]

Horowitz's pushback is the sharpest moment: AI can patch bugs, but we have "50 years of code and architectural ideas that weren't built for this world," plus massive centralized honeypots of consumer data — he asks whether centralized data architecture is even viable now.

https://t.co/E6vb8ygTwx

[triple_minded @triple3minded]

He warns that intelligence will otherwise become a class-division tool, and invokes Marx: capital previously needed labor to create value, but capital plus AI plus robots does not, accelerating inequality toward a Ready Player One or Matrix-like split.

https://t.co/A3bCQzgurr

[triple_minded @triple3minded]

Earnings thesis

Analysts usually cut EPS estimates as earnings season approaches, but unusually raised them ahead of both Q2 and Q3 2026.

Despite the higher Q2 bar, 87% of S&P 500 companies reported positive EPS surprises, reportedly a record-high breadth of beats.

Over the past 20 years, the quarterly beat rate never dropped below 63%.

The typical size of an earnings beat remains modest—around 5%.

Therefore, estimate revisions may reflect analysts' calibration process as much as corporate performance: "estimates miss earnings," rather than earnings missing estimates.

[triple_minded @triple3minded]

Wisdom Heard:

People move the goal posts on you to build a position that's asymmetric in their favour. Asymmetric for them usually means asymmetric against you.

When the world turns in a bad way, the hard part isn't the unfairness — it's that you can't predict when someone will turn and move the goal posts against you. At some point the behaviour stops being analysable.

Step back, redirect your attention, protect yourself by minimising the touching surface with non-analysable people first.

[Big Cheesecake @BigCheeseCake24]

@kristelfung Andy Grove, ex-CEO of Intel, “It is important to say 'no' earlier rather than later... Remember too that your time is your one finite resource, and when you say 'yes' to one thing you are inevitably saying 'no' to another".

[triple_minded @triple3minded]

So well said. It fits the incentives so well. Dario is addressing his own guilt as well.

[Still True Secret @stilltruesecret]

@chamath @bgurley I think it is talent management. I guess Dario is addressing his team on any potential (philosophical) doubts and guilts arisen from their tech breakthroughs.

[triple_minded @triple3minded]

🙏🏻

[Still True Secret @stilltruesecret]

@naval Share here some great words I heard: “I will counter him, the truth is just plain and raw 😂”

[triple_minded @triple3minded]

China's top AI leaders held an OpenClaw open-source roundtable on Mar 27. Zhipu CEO Zhang Peng called OpenClaw "scaffolding" for building agent apps on foundation models — and announced GLM-5.1 + GLM Coding Plan. Xiaomi MiMo's Luo Fuli said OpenClaw's open-source design raises the floor for base models while extending the ceiling. Infinigence CEO Xia Lixue is building infra to make token factories more efficient for Zhipu, Kimi, and MiMo. The consensus: AI agents are the next battleground, and open-source frameworks will define the ecosystem.

[triple_minded @triple3minded]

Cyberstarts founder Gili Raanan notes: Israel produces ~350-400 new cybersecurity teams per year, but only 1-2 out of 150 become unicorns. Seed valuations have risen sharply while feedback cycles remain brutally slow. Cyberstarts created an "employee liquidity fund" underwriting annual tender offers — a talent-retention tool giving employees public-market-like liquidity while staying private.

https://t.co/JKyPB7Km0u

[triple_minded @triple3minded]

SenseTime (https://t.co/8yzMxMJM9u) CEO says their new NEO model (multi-modal architecture) hit 94.3% on complex agent task completion using OpenClaw benchmarks — vs Claude 4.6's 50% — with lower token costs. The argument is strongest in Chinese-language and vision-heavy office workflows. SenseTime's total compute = ~40K PetaFlops; ~6,000 PetaFlops is domestic Chinese hardware (~15%).

https://t.co/bRtxLc1mQW