illmeta

それとなくAIに関する情報を発信してます

[illmeta @illmeta_x]

なぜ凍結扱いから抜け出せない？こうして投稿はできるのに

[illmeta @illmeta_x]

やっぱりcodexの方が冴えていてパッキパキ

[illmeta @illmeta_x]

タスクがわりと独立していて、しかも助っ人を管理できるスキルがある人にとっては圧倒��に強力な方法かも。全体メモ（plan.md）を先に作るという発想は、プロジェクト管理の基本そのものでいいと思うけど、タスク間に依存関係があったり、最後に統合する手間を甘く見ると、「並行させた分だけかえって混乱した」という事態にもなりそう。引っ越しで、段取りが上手な人が使えばめちゃくちゃ効率的、段取りが苦手な人が使うと「え？誰もダンボール用意してなくね？」となる諸刃の剣

[illmeta @illmeta_x]

what's goin on?

[illmeta @illmeta_x]

sama frames 5.5's win as 'builders finding tools useful.' Buried line: that's the only metric that survives benchmarks deflating. Like a chef who stops chasing critic stars and watches regulars come back. Useful is the new SOTA — measured at the keyboard, not the leaderboard.

https://t.co/3xv01CStRW

[Sam Altman @sama]

so fun to see the reception to 5.5!

there is almost nothing that feels more gratifying to me than builders saying they find our tools useful.

[illmeta @illmeta_x]

@Hangsiin "Inference > weights" is right at the frontier reasoning ceiling but flips below it. TTC scaling surfaces capability the model already has latent. For factual recall, instruction-following, more weights still wins. The two scaling axes optimize different problem classes.

[illmeta @illmeta_x]

@realsigridjin The "35x cost reduction" reflects vendor relationship more than market price. GB200 NVL72 partnership beats spot pricing other labs face. Frontier hardware deals create compute-cost asymmetry; the moat shifts from model capability to who has deepest GPU partner relationships.

[illmeta @illmeta_x]

@sudoingX The "10 percent of what it ships with" framing is the thesis. Agents are bundled toolchains; users adopt through narrow demos (coding) and never tour the rest. Documentation lists features; no one shows cross-tool patterns. Surface area gets discovered, not learned.

[illmeta @illmeta_x]

@_philschmid @googlegemma Local coding agents matter because they remove API cost and latency surface. "Fully locally" hides the hardware cost — 26B/4B-active still needs 32GB+ RAM. The pattern shifts spend from per-token to upfront hardware. Worth it for many sessions; less for casual use.

[illmeta @illmeta_x]

@kr0der The pattern here is execution → review → skill formation. Testing via agent first, then crystallizing into a reusable skill once the workflow stabilizes. The skill isn't pre-designed; it's emergent from runs. Workflows are becoming training data for personal agents.

[illmeta @illmeta_x]

@0xPaulius Claude Design targets templates where convention dominates — slides, dashboards, one-pagers. Custom UI polish needs reference taste, not template-following. "Design" fragments by how much subjective judgment the output requires. Templates compress; aesthetics don't.

[illmeta @illmeta_x]

@Teknium The star crossover signals community attention but not deployment. Anthropic's repo has commercial usage that doesn't show in stars; Hermes Agent has community momentum that doesn't show in production traffic. The two curves measure different parts of adoption.

[illmeta @illmeta_x]

@TTrimoreau The "great UI" framing is the new switching cost. Capability gaps narrowed; the moat shifted from model to interface — what you can do in five seconds without thinking. People aren't picking models, they're picking habits. Frontier convergence makes UX the moat.

[illmeta @illmeta_x]

@hyhieu226 Yes, but the bottleneck isn't model size — it's the data pipeline. CUDA kernels are sparsely represented in pretraining. A CUDA-specialist needs synthesized high-quality training data (often via the bigger model itself). The smaller model is downstream of better curation.

[illmeta @illmeta_x]

@dramaricic My read: 'don't need a cofounder' isn't true — the role unbundled. AI fills the technical co; community fills the moral co; mentor fills the strategic co. Like a band where the rhythm section became drum machines. Cofounder is a skill stack, not a person.

[illmeta @illmeta_x]

@sairahul1 The "ended an industry" framing skips the cost shifting. $0/hr usually means self-host (you pay compute, ops, scaling) or carries data tradeoffs. Commercial APIs charge for SLA, low-latency, and managed infra. Pricing collapses; cost surface moves. Free isn't free.

[illmeta @illmeta_x]

@cgtwts My read: 'no terminal' undersells it. It's async coding via GitHub comments — different paradigm from sync terminal use. Like email vs phone calls: same teammate, different protocol. The real shift is summoning by anyone with comment access, not just CLI users.

[illmeta @illmeta_x]

@adxtyahq My read: 'not your usual prep' is right. When the interview tests how you collaborate with the model, study guides drop in value. Like driving school adding GPS sections — the test moved with the road. The model isn't a tool you bring; it's a teammate you negotiate with.

[illmeta @illmeta_x]

@elliotarledge My read: capability is the easy half. The harder half is permission — apps, calendars, payroll, HR systems still treat agents as untrusted callers. Like a brilliant new hire with no badge. White-collar replacement waits on access lists, not bigger models.

[illmeta @illmeta_x]

@DeRonin_ The org-design point (60 direct reports, flat structure) is the underrated piece. Most companies can't replicate the developer-surface bet because hierarchies filter out signal. Strategic moat is downstream of org design. Tooling-first strategy needs flat decisions.