gobdovan

There are recorded cases of real regressions, but they're better characterised as incidents, not nerfs, e.g.: https://www.anthropic.com/engineering/april-23-postmortem

Btw, you have a typo in the twitter handle on your profile (not in your comment), 'thesilencesturns'.

Only for the $200 one. The $100 one was already pretty poor value for allowance/$. Without the old $200 sub, I wouldn't have used Codex.

They have also cut allowances for subscriptions in half. So even in the best case scenario it's about 2.5 times cheaper for Codex users. They just seem to have matched Claude Sonnet 5.5 *API pricing*, but from what I see online, it seems Claude Code now has a much more generous subscription allowance.

I was pointing out a minor consideration for Rust on server vs TS on local in general, not current `cf` specifically.

There's also a (maybe more minor but still interesting) explanation. On a server, you know quite precisely your load and you want to be able to causally trace bottlenecks, which is simpler with AOT compiled programs. On users' machines, they may not give you full telemetry and may use your system in quite variable ways. So V8 comes in quite nicely and optimises hot paths on workload it observes on each users system.

I think this is valid now, but not guaranteed to be valid forever. For engineers, there was a period where more checks, more tests, more auto code reviews improved results quite a bit. People were consuming tokens like crazy (including me). Then things improved via better effort/thinking levels, where you could see repeated code reviews plateaued, so now people don't really do that quite as much.

There was also a period where specifically OpenAI models would always have to comment something in code review and the builders were agreeable up to listening to each nitpick. If you'd have a loop of build->review->build->review, it would take maybe 5-7 rounds for it to 'settle' and not find the smallest nitpicks to argue about. Tried it this week with Astra reviewer and it's about 0-2 review loops (never had a LLM accept a change without nitpicking first try before Astra).

There was also a period where you'd have to give quite specific instructions for agents to keep iterating, but now agent are pretty proactive and try to finish tasks you give them unsurprisingly most of the time.

So, while there's a shortcoming of LLM+harness and engineers observe more tokens improve things even logarithmicly, you'll see more tokens seemingly abused by engineers.

> Chess only teaches two general lessons about strategy: look more than one step ahead, and invent heuristics

That is true for an engine.

Meta-games around studying your opponent, specific match preparation and all those game theoretic stuff are so pervasive people claim they are 'killing' chess at top levels. Anyway, for humans, chess at high levels of mastery is a striking display of bootstrapping some insane cognitive machinery around memory, processing speed, 'expanding' working memory via understanding:

[0] Alekhine blindfold simultaneous vs 32 opponents: https://www.blindfoldchess.net/assets/images/alekhine-excerp...

[1] Reingold, Charness, Pomplun & Stampe (2001), "Visual Span in Expert Chess Players": https://pubmed.ncbi.nlm.nih.gov/11294228/

[2] Connors, Burns & Campitelli (2011), "Expertise in Complex Decision Making: The Role of Search in Chess 70 Years After de Groot": https://onlinelibrary.wiley.com/doi/epdf/10.1111/j.1551-6709...

"Can you stack your family?" - Michael Falk, The Onion

Dynamic programming. Although if you don't know the distribution of how the bets will look beforehand, there's no real optimal strategy. Think you see first bet has 99% chance of giving you x2. Do you bet everything and take it? It could either be the worst bet of the 20 bets, in which case you just took 1% chance to lose everything unnecessarily, or it could have been the single bet that would be winning money, in which case betting it all would be the only strategy that could win.

We're in such a bull market, even shortages grow.

I'd refine it as 'If feasible, try fixing your diet before going for pills'. The body is adapted to function within a healthy nutritional range. If your diet falls outside that range, it's expected that your body won't function properly. If problems persist despite a good diet, then pills become a more reasonable option.

Humans are very inefficient when it comes to navigating the web, but also take actions pretty fast when completing forms. You don't really need advanced ML to see bots spend two seconds to read a full page, then spend 10 seconds just to click two buttons a human would click together in under 2 seconds. The amount of sophistication in bot detection peaks at about 'if user searches 20 queries in less than 5 minutes on our search engine and uses incognito, CAPTCHA them'.

Because of this, perfectly mimicking humans is not a good goal for a bot (as it is the case for AI in music), because they would become very inefficient, at least latency wise (throughput could be engineered around by scraping many unrelated webpages in parallel).

You can't reasonably expect generic ASR to infer tmux from "tee-mucks". "tee-em-you-ex" works reliably if you're ok with capitalisation for your use case.

How is anyone making money from courses if dabeaz isn't? He's got the word of mouth going on, celebrity status in the Python world, world-class courses, it doesn't make sense to me. I'm not asking rhetorically, am truly curious about what's going on.

When on Speaker/hands-free mode, it just closes screen. They assume you wouldn't press lock button while against your ear, because it closes the screen automatically. The problem is that there's some bugs that keep the screen open sometimes, or you may use it in a quiet room as if it were on hands-free.

Markets and evolution are both knowledge processes. I think you would enjoy Donald Campbell's essay on evolutionary epistemology from 'The Philosophy of Karl Popper'.

Parent's comment is observing a prerequisite to markets, humans are the only animal with global markets and biological selection happened to produce the kind of creature that can stabilise around markets.

It's also selection that produced the current form of markets themselves. The world could have been otherwise; our current markets are contingent selection products, not analytically necessary institutions, and not guaranteed to be the uniquely best mechanisms for producing abstractions and tools.

Were 8 billion people all working toward more detail? Isn't civilization also the process by which detail is pruned, compressed and made into abstractions small enough for one mind to use?

I usually view it the opposite way from the article's perspective. There's surprisingly little detail we rely on, yet things work out somehow.

There's the Popper observation that any model of reality has zero chance to be true, since our models are finite, yet we're trying to describe a fractal reality. It's amazing how few levels of decomposition we need to go through to get something useful, like the stairs in the article (3 decomposition steps, as compared to thousands). If I were to never interact with reality and rely on pure reason alone, I would expect nothing humans ever do to work.

Abstraction and exploration are unreasonably effective.

Was discussing the content, not the author. Yes, I am challenging vague discourse about echo chambers, since the mechanism behind an echo chamber is selectively listening to some opinions, while dismissing other sources of authority or correction. To see why this is a problem, consider that parent comment is compatible with both:

"Flat Earth is a cancerous echo chamber"

"As I grew older and became better at arguing, I learned to explain why universities teaching that the Earth is spherical are cancerous ideological echo chambers."

> isolating themselves & having their [...] views spread far and wide

> most cancerous developments & the less contentious it becomes

Your comment complains that people cannot articulate their reasons, while making a sweeping, emotionally loaded claim whose reasons are themselves barely articulated.

Relational algebra, EAV model, MVCC, Nix and a few principles around removing data only if you can prove it's reconstructible some other way.

My email's on my profile.

I'm developing a similar project, I also added scripts to it so it works like an hermetic/replayable system too. Do you use yours for anything cool? Maybe a truth maintenance system of sorts? Do the queries get unwieldy at some point?

I heard people with dogs claiming similar effects.

I'd like to imagine that the creator heard the 'Einstein zebra puzzle' (the puzzle about different nationalities having different pets and smoking different brands of cigarettes) and just made the pets have different nationalities and smoking the cigarettes directly.

Actual stealing is an even more impressive skill. Usually involves intensively trained sleight of hand, elaborate ruses, a very good understanding of theory of mind regarding the victim's attention, and planned deescalation paths in case you're caught.

This is a pre-AI phenomena. I observe it quite a lot with stuff I did in high school but usually with complex problems. What's generally happening is that you were working with pen and paper through a hard problem. With adult brain, you'd expect just to know the answers, but in reality you're not much smarter than you were at 14, so you need to do the thing properly.

Also if you help little kids with homework, you'll see that some problems are quite difficult as well and require you to actually think, even if it's problems for 10 year olds.

You can go down or right at any point. To go in bottom-right corner, you need n down steps and n right steps. In how many ways can you arrange n things on type A and n things of type B? In C(2n, n). The problem is about modeling, once you model it correctly, you get the definition of combinations.

I'm not sure I'm following you. What does 'duplicate Google Docs' mean? We already have OSS rich text editor (ProseMirror) and CRDT implementations (yjs, automerge). Are you talking about replicating every single enterprise Docs has? Integration with a large suite of other services you control? Replicating the business model too?

'duplicate Google Docs' spans everything from providing the actual service to replicating the world and becoming Google.

An option won't fire you 1 day before cliff.

> Style is not neutral; it gives moral directions.

> Nowadays every business in America says how warm it is and how much it cares — loan companies, supermarkets, hamburger chains.

Guess which one is AI and which one is a quote from Martin Amis.