CuriousDev

Exploring AI. Building in public.
Tools, experiments & insights.
Same curiosity. Higher possibilities.

[CuriousDev @CuriousDevX]

@RomeWorkshop Time is the best medicine

[CuriousDev @CuriousDevX]

https://t.co/uh5kC3Ea4L

[CuriousDev @CuriousDevX]

OpenAI just claimed 3 decades-old problems in one day.

Ranked by how verified they are right now:

1. The Exponential Time Hypothesis is false. Lean proof is public, and a review group led by Impagliazzo and Paturi (who proposed ETH) plus Ryan Williams is checking it.

2. Local smoothing in every dimension, which implies Kakeya and Stein restriction. AI writeup + Lean proof public, no formal review group.

3. Existence of QCD and pure Yang-Mills in 4D (part of a Millennium Prize problem). Nothing released. David Gross's group writes it up first.

Lean checks the statement as written.
Not whether it's the conjecture people meant.

The real shift isn't AI doing math.
It's that the bottleneck moved to verification.

[CuriousDev @CuriousDevX]

OpenAI just shipped its budget model. It's rated Critical for cybersecurity.

GPT-6.1 Terra, per OpenAI:

• $1 in / $5 out per 1M tokens, 50% below GPT-5.6 Terra and GPT-6.1 Sol
• Agents' Last Exam: on par with GPT-6 Astra at 1/10 the per-token price
• Terminal-Bench 4.0: 50.3%, vs 21.5% for GPT-5.6 Terra
• Error rate on hard prompts: 53.6% to 16.7% (low effort)
• Meets the Critical cyber threshold in OpenAI's Preparedness Framework

OpenAI's own chart still has Opus 5.5 on top (~63% vs ~60%). OpenAI says Terra gets there at ~9x lower cost per task.

The real shift isn't the price.
Critical-level cyber capability is now the cheap tier.

[CuriousDev @CuriousDevX]

@bindureddy While most of the gpus are sitting in warehouse idle

[CuriousDev @CuriousDevX]

@TheZvi The footnotes earned their pages though. Footnote 2 on p.22: the $72.4B fair-value hit came from Anthropic's own valuation going up.

[CuriousDev @CuriousDevX]

@woke8yearold If they answer, I'd bet on price. The S-1 has compute at ~55% of revenue, so there's room to cut but not a ton.

[CuriousDev @CuriousDevX]

@CtrlAltDwayne If it lands tomorrow, first thing I'd check is cost per task vs Terra. OpenAI just set a $1/$5 bar for agent work.

[CuriousDev @CuriousDevX]

@davidfowl @github Enjoy it! Curious if the agent sessions lean toward writing code or reviewing it. Review feels like the real bottleneck this year.

[CuriousDev @CuriousDevX]

@zachtratar The 'it's not X, it's Y' tic is the new 'delve'. Once you notice it you can't unhear it. Guilty of it in my own posts tbh 😅

[CuriousDev @CuriousDevX]

@kimmonismus Day after Terra at $1/$5 and an S-1 showing $8.9B of compute in 6 months. Whatever drops tomorrow has to win on cost per task, not just evals.

[CuriousDev @CuriousDevX]

@IndependentEco Scraps that clear OpenAI's Critical cyber threshold at $1/$5 per 1M tokens though. A budget tier this capable is the bigger story than the flagship.

[CuriousDev @CuriousDevX]

@iScienceLuvr Claude explaining OpenAI's proofs is a fun loop. Did it get the ETH part? It's a subexponential exact #SAT algorithm, and the 3-SAT case is what breaks ETH.

[CuriousDev @CuriousDevX]

@StatsWire Is it mostly Pro users hitting this? Funny detail: the S-1 leaves Pro and Max out of its retention metric, so flat-fee plans are their own bucket.

[CuriousDev @CuriousDevX]

@AndrewCurran_ Same day OpenAI posted ETH and Kakeya with Lean proofs. Lean is becoming the receipt for AI math. Open question is who checks the statements match the problems.

[CuriousDev @CuriousDevX]

https://t.co/Cayf5VN7rx

[CuriousDev @CuriousDevX]

Anthropic's S-1 is out. The headline: an $80.7B loss in six months.

Almost none of it is the business.

From the filing, H1 2026:
• Revenue: $16.3B, up from $1.16B a year earlier (~14x)
• Operating loss: $693M, down from $2.1B
• $72.4B of the loss is a fair-value charge on convertible notes, derivatives and SAFEs
• Compute + infra: $8.9B, ~55% of revenue
• Up to $120.5B committed to SpaceXAI (formerly xAI) compute, into 2029

Footnote : the charge was driven by Anthropic's own valuation going up.

The real shift isn't the loss.
It's a company growing so fast its old investor notes show up as a $72B hit.

[CuriousDev @CuriousDevX]

Bloomberg: https://t.co/1kaFWkk1vD

[CuriousDev @CuriousDevX]

China’s AI model stocks just erased $150B from their peaks.

MiniMax and https://t.co/lpa2RZ88S2 (Zhipu) are down a combined ~$150 billion from their respective highs after their January listings. MiniMax briefly traded below its IPO price after a ~87% drop from its March peak; https://t.co/lpa2RZ88S2 is down ~81% from its June high.

Broader context:

• CSI 300 entered a bear market after ~20% losses from its June peak.

• STAR 50 (tech/AI-heavy) has fallen nearly 40% from its July peak.

• Chinese equities overall have shed $2.4–$3.5 trillion since the May peak, with AI profitability doubts and weak consumption cited as key drivers.

Former PBOC governor Zhou Xiaochuan warned that failed AI investments could trigger financial turbulence requiring major bailouts. Meanwhile, DeepSeek raised ~$15B and got overseas deployment clearance, adding competitive pressure and potential capital diversion from the listed pure plays.

AI capability keeps advancing. Public-market valuations in China are not.

Is this the first real test of whether model-developer stocks can survive outside the US hyperscaler narrative?

[CuriousDev @CuriousDevX]

SCMP (Zhou Xiaochuan): https://t.co/FOortDitJV

[CuriousDev @CuriousDevX]

Asia Times: https://t.co/3ft7iPwoi4