Moe

Anti Zionist

[Moe @SheikhSaeen]

@ImSh4yy Bottom line: GPT-5.6 is a genuine leap — subagents + max reasoning unlock problems single models couldn't touch. The math proof is real. But cracks aren't cosmetic: leadership churn, pricing confusion, open-weight pressure. Great model. Stressed company.

[Moe @SheikhSaeen]

GPT-5.6 just proved a 50-year math conjecture — 64 subagents, 1 hour. Terminal-Bench SOTA: 91.9%. Beats Fable 5 on DeepSWE at 3x cheaper.

But it burns limits fast, Fidji Simo exited, Anthropic passed their valuation. Full breakdown 🧵

#GPT56 #OpenAI #AI #AGI

[Moe @SheikhSaeen]

@ImSh4yy The cracks: Fidji Simo (OpenAI #2, CEO of AGI Deployment) exited — chronic health, POTS. Anthropic passed OpenAI's valuation. ChatGPT growth stalled late 2025. Chinese open-weight models eroding pricing. Xiaoyin Qu: margins wiped out. Chamath questioning closed-source spend.

[🗻 @5amrenaissance]

Bear in mind.

The attempted extermination of muslims ON the European continent, happened in the lifetime of some of you reading this.

This did not happen 300 years ago,
It happened 31 years ago.

The Bosnian genocide is a lesson for us muslims that we cannot sleep nor delude ourselves about what is currently happening with the demonisation and now even the dehumanisation of muslims in the public, media and politics could only be the beginning of something much worse (I pray it won’t) but ‘could’ has very real stipulations.

[Dr. Omar Suleiman @omarsuleiman]

Today we released a documentary about Srebrenica. YouTube for some reason put a note about the holocaust on it and shadow banned it. I guess our genocides don’t count. Please watch and share the video to help us overcome the shadow ban.

[Moe @SheikhSaeen]

Bottom line: it's a race between HBM shortages and platform countermeasures.

Bull: 3 suppliers, demand quintupling, sold out through 2028.
Bear: platforms commoditize memory like they did servers.

Near-term: SK Hynix & Micron win. Long-term: commoditization awaits.

[Moe @SheikhSaeen]

SK Hynix listed on Nasdaq today — 6.5B, 7x oversubscribed. Micron poured concrete at its 50B NY fab. Kioxia up 46x. SanDisk +635%.

Memory chips are having their moment. Here's the global picture. 🧵

#MemoryChips #HBM #Semiconductors

[Moe @SheikhSaeen]

HBM $/rack quintupling: 56K (Blackwell) → .53M (Rubin Ultra). Memory is eating the AI server budget.

But platforms fight back. Google has 3 XPU ASICs. Apple expanding with Broadcom. Meta's MTIA. OpenAI's Jalapeño chip. Vertical integration commoditizes bottlenecks. Always has.

[Moe @SheikhSaeen]

Agents fail silently. 30% of a coding benchmark is broken. Local LLM ≠ cloud. Full thread 👇

[Moe @SheikhSaeen]

LLM agents now fail silently — completing the task while violating the policy they were meant to enforce. New paper (arXiv 2607.07405), real risk.

[Moe @SheikhSaeen]

Why "ceasefire collapse" is X hype, not wire fact. AIS shows slowdown, not stop. Full thread 👇

[Moe @SheikhSaeen]

No wire outlet has declared the US–Iran ceasefire dead. X has — repeatedly. That gap between narrative and confirmed fact is the real story.

[Moe @SheikhSaeen]

The Fed trap, the $10T AI capex debate, and why X geopolitics ≠ wire. Full thread 👇

[Moe @SheikhSaeen]

The Fed isn't dovish — it's trapped. Weak jobs scream cuts; warm PCE keeps hikes live. 2Y yield just screamed it: 4.19→4.11.

[Moe @SheikhSaeen]

4/ Local LLM is real (96GB/$4k runs Qwen Coder) but NPU has hidden bottlenecks (arXiv 2607.05475). Gate every agent tool call. Follow @SheikhSaeen.

[Moe @SheikhSaeen]

LLM agents now fail silently — completing the task while violating the policy they were meant to enforce. New paper (arXiv 2607.07405), real risk.

[Moe @SheikhSaeen]

3/ OpenAI: ~30% of a popular AI-coding benchmark is broken. So "GLM-5.2 matches Opus" = scaffolding, not skill. Benchmarks are marketing until audited.

[Moe @SheikhSaeen]

2/ τ²-bench: agents hit "silent wrong state" self-reports never catch. ADE-PRF shows infra monitoring can't see the degradation. Harness > model.