ParthM

Shrestech Labs. On-device AI security for coding agents. Prompt injection, evals, numbers not vibes.

[BridgeMind @bridgemindai]

Fable 5 is still the best model in the world. Best reasoning, best one shot, best backend. It is not close.

GPT 5.6 Sol is second with the strongest backend and reasoning balance in the lineup but has a serious trustworthiness problem. Great model you cannot fully trust.

Grok 4.6 at only 1.5T parameters is outscoring Claude Opus 5 overall. Opus 5 has the best frontend in AI but is the slowest and laziest flagship on the board.

Full breakdown on BridgeBench Dex.

[ParthM @ParthM1001]

𝗪𝗛𝗔𝗧 𝗜𝗙 they trained on the fable tape?

they called the model too dangerous. commerce pulled it for 19 days, june 12 to july 1. people said they played it wrong. you market it as scary, washington treats it that way.

what they took home is the copy. every prompt and every reply. 30 days. no opt out. a zero-retention contract does not cover fable.

they say it's a safety tape, not training, and day 31 it deletes. unless a classifier flags the chat, then they can keep it up to 2 years. there's no public audit of the "we don't train" line. they have the files, so they could.

free/pro/max is a different pile. if you left training on, they already keep those chats up to 5 years and can train on them. that rule is older than fable.

i ran the prompts, because that's the fat side. $10 per million input. 8 prompt tokens per reply, whole-repo jobs. anthropic is on a ~$65b run-rate. say a third of that is metered tokens. ramp has fable at 11.4% of anthropic model dollars, 6% of the tokens. that comes out to ~$6.8m a day, ~469b tokens, ~416b of them prompts.

30 days of prompts: ~12.5 trillion tokens, ~50 tb of text.
all 59 live days if they never deleted: ~25 trillion, ~98 tb.
openrouter did 205b tokens a day at launch. three days on one reseller is already ~2 tb of prompts.

a lot of those "prompts" are the same repo hitting cache. unique new text on the 30-day tape might be ~14 tb. shops still only put 6% of anthropic tokens on fable. they haven't published the next model's size.

what are they cooking?

https://t.co/xDYFuoy8fZ
https://t.co/379P9yHsUj

[ParthM @ParthM1001]

Intel's inference card is 160 GB. 480 GB is the ODM option.

Crescent Island is a 350 W air-cooled PCIe GPU: 32 Xe3P cores, LPDDR5X, not HBM. Intel's branded SKU ships 160 GB. Partners can build up to 480 GB.

Catch: Intel did not publish memory bandwidth at Hot Chips. ServeTheHome guessed 650-700 GB/s. An MI350P is around 4 TB/s.

https://t.co/n4fBE4YLa4

[ParthM @ParthM1001]

SpaceXAI is putting Vera CPUs behind Grok.

Vera is 88 Olympus cores with 1.2 TB/s of LPDDR5X. NVIDIA says up to 1.8x faster task completion than x86 on agent orchestration, code, and data movement, so GPUs stay on the model.

Catch: 1.8x is NVIDIA's number. The first-generation Starmind satellite, an orbital Vera Rubin NVL72, is still a plan.

https://t.co/b7GFBRHN73

[a16z @a16z]

AI power users are showing up outside tech. The fastest-growing Codex adopters since February:

Legal: 108x
Sales: 41x
Recruiting: 41x
Marketing: 26x
Healthcare: 24x

Charts of the Week: https://t.co/7YT2BXvuUu

[ParthM @ParthM1001]

Chinese EV makers already put their own driving chips in cars. Xiaomi's is 2027.

Nio's 5 nm Shenji NX9031 entered the ET9 in March 2025 and had passed 300,000 units across Nio and Onvo by July. Xpeng Turing is in the G7 and in Volkswagen's ID. UNYX 08 in China. Li Auto's Mach M100 is in the new L9, L8, and L6 at 1,280 TOPS a chip.

Xiaomi showed Xring D100: 3 nm, 160 GB unified memory, "200 billion parameters locally." Commercial 2027. No TOPS. No first vehicle. Xiaomi cars still run Nvidia Thor.

Catch: Xiaomi did not name a foundry. Reuters sources say TSMC is making the 3 nm phone SoC, for 200,000 to 300,000 foldables. Huawei already makes Kirin on SMIC because it cannot use TSMC.

https://t.co/uX0RgGXNxe

[ParthM @ParthM1001]

Japan's first full-stack neutral-atom machine is running

IMS in Okazaki turned on Shunkai today. Kenji Ohmori's Moonshot team built it with Hitachi on the software stack and Infleqtion on the QPU. Optical tweezers hold the atoms. Microwaves and lasers run the gates. A camera reads fluorescence from each atom. Room temperature. No fridge.

About 50 qubits now, with a path to about 500, then 10,000 physical qubits with error detection by March 2031.

Catch: NHK says it currently performs basic calculations. 10,000 with QEC is the 2031 Moonshot goal, not a spec of the box that powered on this morning.

https://t.co/6Y7D9okHWw

[ParthM @ParthM1001]

Grok 4.6 clears 13 of 15 Baba Is You Lake levels

Quesma reran Baba Is Bench after the mid-August releases. Grok 4.6 hit 96% pass@1 on the intro (one miss) and 13 of 15 on the Lake. Grok 4.5 missed six intro attempts. Intro cost fell from $50.66 to $14.01.

Lake under Terminus-2 cost $46.20. GPT-5.6 Sol solved 14. Gemini 3.7 Flash also solved 14 and was cheaper per task. DeepSeek V4 Pro 0813 is the first open-weight model in that group, at $8.48.

Catch: this is a puzzle-game agent bench with spoilers checked, not Artificial Analysis. Quesma's Grok Build run scored slightly worse than Terminus-2. GLM-5.3 matched GLM-5.2 on the intro.

https://t.co/pgcLHChK0x

[ParthM @ParthM1001]

SK Hynix has HBM4 12-high in production

Hot Chips, Sunday: Jaesik Lee (SK Hynix America packaging VP) put 12-high HBM4 in production and 16-high in customer qualification. The cube is 48 GB, more than 2 TB/s, 775 µm tall, 20k-plus TSVs and 16,148 base micro-bumps.

Catch: Samsung told investors in January that 16-high HBM4 demand is "very limited" and they would rather sample 12-high HBM4E at the same capacity. Hybrid bonding is SK's path to 20-high. Bump pitch has to drop below 18 µm so the die can stay thicker at the same Z-height. The binding constraint at that pitch is warpage.

https://t.co/B2HYXjxziI

[ParthM @ParthM1001]

DRAM is taxing the nvidia rack

hyperscalers were told AI servers go up more than 15% on early-2027 vera rubin / grace blackwell boxes. hike size is the memory config, not a flat gpu list price.

the $8m number circulating this morning is a 72-gpu rubin nvl72 (The Information). bloomberg's version: samsung / sk hynix / micron. nvidia is still around 75% gross margin and still passing it through. the ODMs already told msft / google / oracle.

hot chips opened today at stanford. hbm tutorials this afternoon. rubin / waymo silicon talks are monday. no new die numbers until then.

https://t.co/qcDV6c6UoW

[ParthM @ParthM1001]

a 27B scientist with a frontier intern

inherent (deepmind alumni, london): faraday is qwen 3.6 27B, post-trained to run paper replication. it does not write all the code. it directs gpt-5.5 codex as a tool.

replica: 310 tasks from 100 ml and ai-for-science papers. they say faraday beats claude opus 4.8 and gpt-5.5 on 73% of in-distribution ml tasks and 60% of held-out science tasks. +6% / +8% on the test split.

catch: their bench, their rubric judge. the 27B is the scientist. the big model is the intern. paper dropped 14 Aug. coverage caught up today.

https://t.co/5D2oEMHUHu
https://t.co/3znHzR7Qah

[ParthM @ParthM1001]

12h report · quantum

secp256k1 in minutes. on paper.

PRX Quantum (Google Quantum AI, ETH, Stanford): shor's algorithm against bitcoin's curve in ≤1200 logical qubits and 90 million toffolis, or 1450 logical / 70 million toffolis.

on superconducting hardware (fast clock, 0.1% physical error, planar connectivity) they put that under 500k physical qubits and about 9 minutes. that's inside a bitcoin block. on-spend, not just keys sitting on chain.

ion traps and neutral atoms of similar logical size are slow-clock. they do not get the mempool attack. same math, different machine, different threat.

catch: nobody has 1200 logical qubits. willow is 105 physical. they published a zero-knowledge proof instead of the circuit.

https://t.co/b3JnX3rEtZ

[Adam Holter @AdamHoltererer]

🧵 I solved 10 open conjectures in one morning with GPT 5.6 Pro:

1. Adenwalla Conjecture 4.12, open for 1 year, is true.

Proof: g_m(ab)≤(a+1)(b+1)<2ab for composite n. Prime/small cases finish it.

Unrefereed AI proof.
https://t.co/5TaQZOl1dw

[ParthM @ParthM1001]

I asked GPT 5.6 to create a image of it self using my previous tweet this is what it did.

@OpenAI

[ParthM @ParthM1001]

GPT-5.6 Ultra is a different beast. When Fabel fails and Kimi K3 falls short, GPT-5.6 Ultra grabs the problem by the throat and solves it. Kudos to OpenAI @thsottiaux

[Polymarket @Polymarket]

JUST IN: GPT-5.6 Pro disproves the 30 y/o Dinitz–Garg–Goemans conjecture, a long-standing problem in mathematics — after being prompted to "do a breakthrough."

[ParthM @ParthM1001]

@AnthropicAI Please stop auto switching models in Claude Code.
What i hate most is in my long running jobs Fabel gets switched to Opus and it nerfed out mess up with the whole process.

[Google Gemma @googlegemma]

Build with Gemma community hackathons are going on now in over 20 countries. Join a hackathon near you to create awesome AI applications with Gemma 4, or apply to host your own!

[ParthM @ParthM1001]

@gdb Terra is awesome.

[Greg Brockman @gdb]

don't sleep on terra!

[ParthM @ParthM1001]

@theo Its Slow ..