7777777phil

OpenAI is doing the Joel Spolsky move, commoditize the layer above yours. Google did it with Kubernetes back when GCP was losing on infra. If they don't put a spec out, every orchestrator gets built around MCP or A2A by default and OpenAI is the third wheel. The spec is free, the runtime is where the money is..

AlphaZero worked because chess and Go have terminal rewards and positions you can prove are right or wrong. General intelligence has neither, and the leap from self-play in a well-defined game to self-play in arbitrary environments is the hard part Silver isn't really demoing. Sara Hooker's stuff on scaling laws lines up here (1)

(1) https://philippdubach.com/posts/the-most-expensive-assumptio...

50% cache savings is nice buuuuut.. token cost is the product now, not the model. Every coding agent is converging on the same trick, cache the context across calls. Anthropic's prompt caching, then Cursor's snapshots, now this Reddit thing.

Captive aftermarkets are roughly the biggest hidden cross-subsidy in consumer goods imo.. printer and tractor OEMs price the unit near cost and pull the lifetime margin out of parts, service, and locked firmware. That's why right-to-repair is one of the few issues where farm states and urban progressives end up in the same column.

The veto doesn't read as pro-growth, more "don't apply the brake too broadly".. ME-307 was the first state-wide moratorium to get serious traction and one Jay carve-out was the only thing keeping it from law. NY is queued up next. Grid pushback is the part of the AI capex story the 10-Ks aren't really pricing in: https://philippdubach.com/posts/ai-capex-arms-race-who-blink...

> The model is available today under the Apache 2.0 license on Hugging Face (opens in a new window) and Github (opens in a new window).

Bringing back the Open to OpenAI..

> whether the final result is actually better, or whether it is just a more polished hallucinatio

Agents sampled from the same base model agreeing with each other isn't validation, it's correlation. Cheaper orchestration mostly amplifies whatever bias the model already has. Neat hack though.

In 2018 the Swiss voted if cows are allowed to have horns (1). It was called the horned cow initiative.

(1) https://www.admin.ch/en/horned-cow-initiative

MFN clauses are common in retail, what's different about Amazon's is the enforcement. Manufacturer MAP policies threaten "we won't ship you more," which you can recover from (unlike a listing demotion)

>Generation became cheap, validation didn't this is basically the whole post imo. Also maps to why productivity hasn't really moved despite 93% adoption.. the oversight bandwidth eats the generation gains.

Jevons for code is right but the bottleneck is review, not typing: https://philippdubach.com/posts/93-of-developers-use-ai-codi...

KL(P||Q) penalizes Q heavily when it assigns low probability to things P considers likely, but barely cares when Q wastes probability on rare events. That's why KL regularization in RLHF pushes models toward typical, average-sounding outputs..

Auto-switching across model providers basically concedes the model layer is commodity, which I think is right (1)

tbd whether the skill registry develops network effects or just stays a flat directory. Portable skills as APIs tracks with the broader pattern of agent stacks decomposing into specialized swappable layers, where the defensible asset is whatever process knowledge orgs encode, not the deployment infra.

(1) wrote about it here from an enterprise perspective: https://philippdubach.com/posts/dont-go-monolithic-the-agent...

This sentiment will probably resonate with a lot of people here. I literally won’t use a service if they try to force me onto their app..

cool idea, how far back (in time) do those 27k commits go?

Just thinking how this could maybe used for (automated) research / visualization on the evolution of (spanish - in this case) law

A quick reminder of this gem: This Video Has 74,991,092 Views (https://www.youtube.com/watch?v=BxV14h0kFs0)

The amendment 34 (by Markéta Gregorová, Greens/EFA) to the Sippel Report A10-0040/2026 significantly restricts the ePrivacy derogation (chat control extension until 2027).

It replaces Art. 3 para. 1 lit. a of Regulation (EU) 2021/1232: Processing (scanning) may only be - strictly necessary for technologies to detect/remove known CSAM material (hashes, no unknown content), - proportionate, - limited to necessary technologies and content data.

Today I learned that Gemini can now natively embed video..

Cool Project, thanks for sharing!

The sims are really well done, the dynasty simulator especially. You can actually stress-test the argument instead of just nodding along. Appreciate the craft.

I have issues with the economics though. The income model is calibrated from three separate literatures that were never estimated together. Different samples, different decades, different identification strategies. Then the big move, βIQ drops to 0.10, βW jumps to 0.65, gets asserted as a scenario and fed into the simulator like it’s an empirical result. The interactivity makes it feel rigorous but you’re mostly just exploring the author’s priors.

The skill premium has survived every automation wave we’ve thrown at it, including ones that felt just as terminal. ATMs didn’t kill bank tellers. US teller count went from ~300k to ~500k between 1970 and 2010 (see Bessen paper), because cheaper branches meant more branches.

The essay waves off Jevons with “human attention is fixed” but US legal spend is ~$400B/yr against ~$100B in estimated unmet need (LSC data). That’s 25% latent demand just sitting there at current prices. I would see that as saturated.

The “27.5% programmer decline” is doing a lot of work. BLS SOC 15-1251 (“computer programmers”) is a narrow legacy bucket that excludes software devs, DevOps, ML engineers, all of which grew. Total software dev employment (15-1252) was up in 2024 vs 2022. Classification artifact, not a labor market signal. And the historical base rate on “this time the bridge closes for good” is… zero. Power loom, ag mechanization, manufacturing to services, analog to digital,etc. each killed the old skill-to-capital channel and built a new one within a generation. You can’t just assert AI is different from all prior GPTs, you have to show the mechanism that prevents a new channel from forming. The essay doesn’t really do that for me.

The assortative mating argument cuts against itself imo. If credentials lose signal value, the institutions where sorting happens (elite unis, professional firms) lose sorting power too. The essay predicts mating shifts to “wealth directly” but… how exactly? Credentials were legible because institutions verified them. Strip the institution and you’d expect noisier matching, not tighter. The Fagereng et al. paper it cites is Norwegian data, which has among the lowest wealth inequality in the OECD. Not obvious that translates.

Again I generally like the writeup, and I think the essay is right that capital returns are pulling away from labor income and AI accelerates it. But “the bridge narrows and the crossing gets harder” is the defensible version. “Closes permanently within a decade” requires believing something unprecedented will happen on a specific timeline..

The margin death spiral she describes (accepting 10-15% fees to cover payroll, then needing more clients at worse rates) is the same thing that killed mid-tier law firms and architecture studios.

The collective model is how freelance dev shops have worked for a decade already tbh.

Agent breaks sandbox, accesses the wrong repo, nobody watching.

This is basically why AI coding productivity stays flat at ~10% despite 93% adoption.. you speed up one narrow thing and create new problems in review and oversight that eat the gains.

This is quite niche, using d²C/dt² > 0 as the trigger is clever. In bond math convexity tells you when small rate changes produce disproportionately large price moves, same signature shows up when an agent hits a retry loop.. costs accelerate way faster than the work justifies. The zero-variance check is the nice complement, a healthy agent has noisy cost-per-call, a stuck one looks mechanically identical every time.

Monolithic agent platforms that try to own everything will lose to composable stacks where you can swap each layer independently. I think the "Boss" orchestrator spawning specialized agents is the right call imo.

Open-source matters here because the defensible part isn't the orchestration UI, it's the accumulated process knowledge each org builds on top. Tried to map out why monolithic agent platforms are a trap: https://philippdubach.com/posts/dont-go-monolithic-the-agent...

> Given the cost the company will incur to install the capacity for dynamic pricing in its stores, it would be corporate malfeasance if they did not believe doing so would not only recoup the cost, but add profit as well.

dynamic pricing and scalping are the same mechanism.. As an economy student I like the efficiency, as a consumer I hate the asymmetry and inevitability

Most ML practitioners use L1/L2 daily without realizing they're making Bayesian prior assumptions. Gaussian prior = Ridge, Laplace prior = Lasso. Once you see it that way, "choosing a regularization strength" is really "choosing how informative your prior is."

Novo already guided for its first revenue decline in modern history (adjusted sales down 5-13% for 2026). Stock is down 75% from June 2024 peak. The problem isn't just generics, Lilly's tirzepatide beat CagriSema in the head-to-head, and Lilly has orforglipron (oral, way cheaper to manufacture) coming. At 11x forward earnings Novo is pricing in catastrophe, buuuuut.. the structural problems are real this time.

so the model underneath is swappable, but the layer that remembers what you asked it last week isn't. Persistent context is the switching cost here. Anthropic just extended that from desktop to mobile.

FY2026 Pentagon's AI budget jumped 7x to $13.4 billion, now larger than Anthropic's annualized revenue. Once you're on an IDIQ contract with classified compute, good luck switching. Security clearance processing alone takes 243 days. Palantir figured this out years ago, 55%+ of revenue from government now.

Tried to summarize some of that in my recent writeup: https://philippdubach.com/posts/when-ai-labs-become-defense-...

my bad, I just saw 2 hours ago on this one - either way they could be merged

Already posted here earlier: https://news.ycombinator.com/item?id=47376119

I run a Hugo blog and I get more interesting referral traffic from Kagi's small web index than from Google at this point. 5,000 curated sites is small enough to be useful most "indie web" directories are graveyards unfortunately..