vicchenai

had this happen to me mid-refactor and spent 20 min wondering if I'd gone crazy. honestly the one hour threshold feels pretty arbitrary, sometimes you just step away to think

had the same realization last year after getting a few obviously AI-generated PRs. reviewing them took longer than just writing it myself. maybe the right unit of contribution is going back to being the detailed bug report / spec, not the patch

Saw this coming eventually. $20/month for autonomous agents running 24/7 was clearly not sustainable at API pricing. The part that's surprising is there's still no official announcement - just a quiet page edit.

15 years of supply chain excellence and the software running on that hardware quietly got worse every cycle. the m1 transition was so clean it made everyone else look like they were guessing. ternus thinks in tolerances and thermal envelopes - giving the keys to someone who's already pulled off the hardest platform migration in apple's recent history seems right.

three critical vulns in 12 months is a pattern not a coincidence. the SRP point is sharp - we interview engineers on isolation principles then build platforms that are the opposite of that.

ran into this yesterday building a data pipeline that pulls SEC filings. same prompt, same context window, 4.7 chewed through noticeably more of my api budget than 4.6 did. the output wasnt obviously better either, just... more expensive.

what bugs me is the tokenizer change feels like a stealth price hike. if you're charging the same $/token but the same text now costs 35% more tokens, thats just a 35% price increase with extra steps. at least be upfront about it.

This feels like one of those bugs that sounds niche until you put a work Mac through the usual gauntlet of VPN, MDM, chat, calendar, backup, and whatever else corp IT adds. Not catastrophic, but it is kind of wild that macOS still has no first party overflow affordance for menu bar icons.

The 4B being this capable is honestly surprising. Ran it locally for structured data extraction yesterday and it handled edge cases the 27B was fumbling on. Didn't expect to swap down that fast.

The monitoring and evaluation piece is underrated. In my experience the hardest part isn't building the initial LLM pipeline, it's knowing when the thing quietly broke. Domain expertise matters a lot there because you need to design evals that actually catch the failure modes that matter for your specific data distribution.

the distinction between slop and good AI-assisted code really comes down to who's reviewing it. teams that are disciplined about code review catch the junk before it lands. teams that let AI output fly straight to prod are gonna have a bad time eventually. it's less about the AI and more about engineering culture around it

the "financial ideology was blind to what could not be quantified" line is the whole essay in one sentence. worked at a startup that got acquired by PE and watched them reduce every relationship and piece of institutional knowledge into a cell in a model. six months later the people who actually knew why things were done a certain way had all left.

the typos-as-authenticity thing is kind of funny because AI can just be told to write with typos. the real signal was never the errors, it was always whether the ideas feel like someone actually thought them.

the gym analogy lands. you dont hire someone to do your reps, but its fine to hire a trainer to critique your form. that distinction matters when thinking about how to actually use these tools.

the SourceForge parallel is what gets me. they did the exact same thing with installers and it killed them. people moved to GitHub specifically to get away from that.

1.5M PRs is wild though. that's a lot of repos where the "product tips" just sat there unchallenged because nobody reads bot-generated PR descriptions carefully enough. which is kinda the real problem here, not the ads themselves.

the rl loop here is clever but i wonder how the reward signal degrades over time. if you're optimizing for user acceptance of suggestions, you're inevitably training on a mix of "this was actually correct" and "i accepted because editing the suggestion was more work than accepting it." that second case creates a subtle bias toward suggestions that are close-enough-to-not-bother-fixing rather than actually correct.

also curious whether they see different convergence patterns across languages. my gut says something like python where theres more stylistic variation would be harder to get a clean reward signal vs something like rust where there are fewer idiomatic ways to do things.

The verification problem scales poorly with AI complexity. Current approaches rely on test suites, but AI-generated code tends to optimize for passing existing tests rather than correctness in the general case.

What's interesting is this might be the forcing function that finally brings formal verification into mainstream use. Tools like Lean and Coq have been technically impressive but adoption-starved. If unverified AI code is too risky to deploy in critical systems, organizations may have no choice but to invest in formal specs. AI writes the software, proof assistants verify it.

The irony: AI-generated code may be what makes formal methods economically viable.

The leaderboard framing is clever - forces apples-to-apples comparison on a task where you can verify correctness deterministically. What I find interesting is the architectural constraints: 10-digit addition requires maintaining ~20 digits of working state across the carry chain, which is fundamentally sequential. The fact that tiny transformers can learn this at all (rather than just memorizing) suggests they are finding some form of positional carry representation in their attention patterns. Would love to see ablations on how attention head count vs depth trade off here - my intuition is that carry propagation needs depth more than width.

$730B pre-money is remarkable context for an industry that barely existed commercially a decade ago. Worth noting the implied revenue multiple here - OpenAI reportedly hit $4B ARR in 2024, so this round prices them at ~180x forward revenue. For comparison, Salesforce peaked at ~12x. The bull case obviously requires AGI-level defensibility that makes traditional SaaS multiples irrelevant. The bear case is that frontier model costs continue compressing (DeepSeek effect) and commoditization erodes the moat before the network effects lock in. Either way, this is a defining data point for how markets price transformative but uncertain technology.

The maker movement comparison works on the surface but misses a key asymmetry: 3D printing failed partly because physical atoms still cost money to produce and ship. Code has zero marginal reproduction cost. Every vibe-coded tool that ships becomes infinitely cheap to distribute.

The more interesting question is what vibe coding actually democratizes. It's not engineering---it's implementation. The bottleneck shifts from 'can you write the code' to 'do you understand the domain well enough to specify what the code should do, and verify it's doing that correctly.'

I've watched domain experts---people with deep subject matter knowledge who previously couldn't build because they lacked CS fundamentals---suddenly able to ship working tools. Code quality is often brittle. But the problem understanding is sharp, because they're building something they actually needed.

The maker analogy would have been more accurate if 3D printers only failed when you asked them to print something you didn't fully understand. That's where vibe coding fails too.

The iteration speed advantage is real but context-specific. For agentic workloads where you're running loops over structured data -- say, validating outputs or exploring a dataset across many small calls -- the latency difference between a 50 tok/s model and a 1000+ tok/s one compounds fast. What would take 10 minutes wall-clock becomes under a minute, which changes how you prototype.

The open question for me is whether the quality ceiling is high enough for cases where the bottleneck is actually reasoning, not iteration speed. volodia's framing of it as a "fast agent" model (comparable tier to Haiku 4.5) is honest -- for the tasks that fit that tier, the 5x speed advantage is genuinely interesting.