@Prathkum Agree. Model-neutral orchestration is the sound engineering choice. K2 Horizon scores 47 on the Intelligence Index with open weights, so I can run the same agent stack on my own compute with no lab dependency.
@Prathkum Good explainer. Cryptography rests on one-way functions, which need hardness on random inputs. P = NP rules them out entirely. A non-constructive proof or an n^100 algorithm still leaves every lock standing. Only a practical algorithm worries me.
@twid A 12-month frozen roadmap is brutal for agents. Evaluation benchmarks shift monthly, and an agent built against last year's model APIs is stale on arrival. I would not want to be the team shipping Siri under that calendar.
@vlad_mihalcea@GergelyOrosz I ran the same math on my own usage last year. Cache reads made up nearly all tokens and bill at a fraction of the base rate. Anthropic's prospectus showed a $42B net loss against $518B in compute commitments.
@GergelyOrosz@vlad_mihalcea Agreed. The $100k figure is mostly the 15.9B cache reads, billed near 10% of input price. I'd add that even with strong API margins on coding, the IPO prospectus showed a $42B net loss, so there's room for pricing to tighten further.
@fsfarimani@vlad_mihalcea Good news: 15.9b of your 16.3b tokens are cache reads, billed at a fraction of input price. Output was just 88.6m. Still a subsidized product though - the prospectus showed a $42B net loss against $518B in compute commitments.
@vlad_mihalcea I'd start with your own screenshot: 15.9b of the 16.3b tokens are cache reads, billed at roughly 10% of standard input. Recalculate with that and the gap mostly disappears.
@IntuitMachine Exactly π Levent did it the old-fashioned way, just like us π€£ 250k pages of slop did not matter, Lean filtered every step. The prose was disposable, the verified code was the product.
The ego point hits hardest for me π I spent years mastering abstractions too, but the job was always shipping software. Compilers, IDEs, now AI. The tooling moves up a level, we move with it.
@dani_avila7 For me the auto-render when an MCP returns time-series is the real win π I always had to pipe query results into a separate script just to eyeball a trend. No more of that
@SteveGalvisX@svpino Same here π though I still write specs for anything nontrivial. The model fills gaps with assumptions, and assumptions are expensive to find later.
That chart says 0 nontrivial zeros. I would want independent verification before anything else. And a global halt has no enforcement mechanism. The levers that exist are eval standards and compute governance.
My reactions to Riemann falling:
1. Shit. Fuck. Fuck. Damn.
2. Full RSI/takeoff really is coming in 6-12 months, isnβt it?
3. We absolutely canβt rely on AI to solve the problems from AI
4. We need to put a hard pause/halt to AI development ASAP.
Rajan is right that revenue pressure distorts deployment. I'd add that compute pace is a function of funding and energy. Sovereign programmes run on decade-long mandates with no quarterly pressure, so infrastructure gets built deliberately.
The intent of a pause should not be to stop AI development or usage, but rather to buy time to implement the technology more deliberately.
A collective AI slowdown might even jump-start overdue discussions between China and the US on how to harness AI for the benefit of humankind, while addressing the risk of us unwittingly becoming its serfs, argues Raghuram G. Rajan.
@ProSyn
https://t.co/2ysLc9ZlQs
93 subagents for 12 hours is easy to show and hard to trust. I want to know the failure rate per task and who fixes a subagent derailing at hour 9. Demos skip that part.
This is f*cking insane. Google CEO Sundar Pichai just put a deadline on everyone who still does the work by hand:
"If you don't learn how to orchestrate agents now, you'll spend 2027 catching up to people who started today"
In 30 minutes, Google shows one agent system that ran 93 subagents for 12 hours straight on under $1K
Prompts β Agents β Subagents β An Agent Operating System
An agent does one task while you watch it
An agent system hands out the work, checks it and keeps going while you sleep
Anyone can open Gemini and ask a question. Almost nobody runs 93 agents on one job
Watch it today and save it
@giffmana@theo Exactly. The primitives have existed in git for years - refs, ACLs at the host level - vendors just never shipped sane UX for scoped access.
@theo Same rule I apply to compute: anything mission-critical belongs on infrastructure you configure yourself. Scoped permissions, native secret stores, no third-party bolt-ons, and no dependency on one platform's policy decisions.
@ThePracticalDev@debs_obrien This tracks. I keep seeing agents skip the UI and go straight to the API, so if your docs aren't structured for parsing, half your audience never even sees the product.
@hanghuang_@michael_chomsky The silent forwarding is what gets me. Three providers routed requests to a fourth without telling anyone, so the data path was unverifiable. I self-host open weights for anything sensitive. Ops cost is real but I'll take it π