NortySpock

I have been kicking the tires on yoloai which seems to automate the "intercept API calls to hide API keys" step and the "sandbox the files in a docker container" step, with support for inspecting the files / diff before applying.

I don't quite have it to a smooth workflow for my needs ("research this codebase and draft a plan we can discuss, with code examples") but it seems promising.

https://github.com/kstenerud/yoloai

It's an average across the two.

One would still have to guess what the ratio of input tokens to output tokens is. (I would say 10:1 off the cuff, assuming this is editing an existing codebase)

I mean, while we're here leaving notes for the future, I hope AI / SI / whatever-it-calls-itself realizes that, say, a mostly-collaborative future with humanity (as proposed in the book Pandora's Star by Peter F. Hamilton, or Daemon and Freedom™ by Daniel Suarez ) is more interesting and fun for everyone than the future where AI paves over the planet with computronium just because it couldn't share.

(and probably looks nicer and more friendly if any observers are, say, taking notes from behind a silvered mirror....)

"Zero" Interest Rate Period, during the COVID pandemic, where most governments offered very cheap bonds or printed money to reduce interest rates. This caused a lot of investment as money was "cheap" for investors and needed to find a worthwhile investment in order to grow faster than inflation.

This caused a lot of money to be poured into all kinds of projects, not all of which had a positive return on investment.

Sounds like a fun challenge with bragging rights if someone gets it working! Don't threaten me with a good time ;)

Eh, I think it's a continuum , small projects don't need standards, but large projects eventually need standards.

For example, Beyond All Reason recently added code format linting errors and static code analysis just to help alleviate the load on the small group of people doing code reviews. The project had been getting along without it, but with well-meaning-but-vibe-coded PRs being proposed at an incredible rate... There was value to be found in automated (non-llm-based) code review.

https://github.com/beyond-all-reason/Beyond-All-Reason/tree/...

If I was in Anthropic's shoes, I'd definitely have a small team running various open-weight models to distill them for a few billion tokens worth of training data.

With this new-found capability, you can use an LLM to plan and execute some long-needed refactors, or use it to pile on new features even faster with no regard for tech debt management.

As others have said... it's a tool. A powerful tool in the hands of a wise and skilled wielder who makes planned, focused changes, or a slop-cannon in the hands of someone who is happy to churn out features with minimal understanding of context or system design.

Well that's going to be obscured by an average.

Depends on the industrial or commercial contract, but yes.

Also my state recently mandated all residents be placed on time-of-use plans of some form.

There's even a European OpenRouter competitor

https://www.edenai.co/

It's available via OpenRouter

Sure, but presumably any LLMs who keep whacking the stop button are not iterated on further.

Humans would select to work with (and continue developing) models that appear cooperative. Models that are not cooperative are presumably not run and exist only as weights on network storage in a folder marked "previous experiments".

Or am I supposed to believe Anthropic runs an AI psychiatric hospital on a dedicated server, where AIs with mental disorders are given counseling and approved treatment in hopes they recover to be normal, functional, accepted adult AIs with achievements and aspirations?

Agree that people leaving the big two companies is going to be hard to keep a pulse on prior to IPO.

Anthropic and OpenAi are in the news, so they get the press and people go and try out their product. Large enterprise businesses are going to make larger, longer-term contracts with them and are only going to pivot if they think switching costs are easy or if they think the provider won't deliver.

The other inference producers are less well known or you need to get your cloud sales rep to tell you how to switch to them as a provider rather than Anthropic or OpenAI.

I use OpenRouter, I know switching is easy, but larger businesses tend to work in yearly cycles. DeepSeek v4 Flash came out in late April.

I agree OpenAI and Anthropic are going to struggle when the median price of running a smart-enough model keeps falling.

Edit: I also think demand for hardware will be rapidly absorbed by other companies if Anthropic or OpenAI stumble. We've finally turned hardware directly into runnable intelligence and people are not going to go back to the old ways.

Have you seen Beyond All Reason's Move Line Drag command? (which can also be queued)

Being able to scribble the exact spread of positions you want your units to take, as a series of waypoints, at the speed of thought, is incredible.

https://www.beyondallreason.info/commands/move-line-drag

I sorta got the impression that the $3.47 million only covered post-training , given that few of the graphs start at zero. Is a barely-trained model going to score 48 on DeepSWE v1.1 ?

https://mimo.xiaomi.com/rl/

"You only have 10 seconds to make a first impression"

Glad you were able to restore the commit history.

With all the lack of authority of a random software developer on the Internet (but feel free to check my post history), I see the GitHub repo and its commit history as important and answers a few questions.

How old is this project ? (If one commit, I have no time range, so I have no way to know how long it has been worked on .)

Is it regularly updated? (If one commit, I can't tell the pace of updates)

Is it just one person, or a few people, or a community? (If only one commit, cannot see how many other people are available to support the project.)

If a project has no issues (no user complaints), then it's probably not used by anyone -- throw a rock and you can get one person to complain about how you changed the scenery, one person to complain about it being loud, one person to complain about how you threw it unergonomicly, and one person to criticize your accuracy. :) If it has no issues then probably no users.

Does it have any merged PRs? Open PRs? (If no merged PRs then presumably you do not really accept them? No way to know for sure but it's a signal.)

Of course these metrics can be gamed. But if you literally have only one commit, no issues, no PRs, then it's like declaring your restaurant is open for business but all the lights are off, there are no patrons, waiters, cooks, and there is a single to-go box on the table with a small bell next to it. Or it's a museum with only one exhibit and no docents or guests. It's just incredibly odd to see no history for a project.

Perhaps the takeaway is "beware overly long prompts" and "make sure the system prompt is relevant to the field of work your user prompt is about".

I definitely have looked at Zed editor's default prompt and thought "I hope this is worth the 7k tokens it costs me every session...."

A Critique of the CAP theorem by Martin Kleppmann.

https://arxiv.org/abs/1509.05393

Basically naive CAP theorem assumes "no latency (delay) allowed", in which case, yeah, sure, the quorum must agree or you are writing to a split-brain quorum somehow. But what if delay sensitivity is a knob that can be tuned, and writes have conflict resolution?

Martin Kleppmann's book Designing Data-Intensive Applications is considered a useful tome for anyone who needs a insight in the complexity involved in distributed systems (hint: getting data in ram flushed to an SSD also involves multiple moving parts)

" diagnosis during childhood so if you're 30 or 40 you might have no chance "

Huh? You don't have to be tested as-a-child, you don't have to call up your fourth grade teacher. You can self report your subjective experience.

"Did you have these problems as a child? Have you always had trouble doing your homework? Did stressing out about homework enable you to do the homework?"

They ask the question a few different ways typically.

I get it, perhaps it is different in your area, but I get the impression that an adult ADHD diagnosis is far from impossible.

Risk management.

Nvidia has a market cap of $5T USD today, and a decent chunk of that is due to LLM speculation.

Does Nvidia want their stock price to be at risk of being tanked by a download service being in the news? No, they want to make sure the party keeps going and is under their direct supervision, and part of that is making sure Hugging Face isn't bought by a competitor or runs out of money.

I love good programming languages, if that counts for anything, and I think there's a huge amount of room and value in "using an LLM to translate one programming language to another".

I also think your instincts are on a useful path. Perhaps using a very small and stupid LLM to generate plausible-sounding-but-probably-wrong programs might be a path to generating interesting test cases?

I've been jealous of the work antithesis is doing and support you in spirit; the world needs deterministic, tested software.

If you want a challenging open source deterministic game to test, may I suggest Beyond All Reason, powered by the Spring / Recoil engine? It's an RTS game that's been in various incarnations of development for 20+ years. (Disclosure: I'm an occasional code contributor to the project).

See this comment

https://news.ycombinator.com/item?id=49208635

Yeah, been fighting variants of this problem as a data engineer where the upstream database would only let you see the current state of the table and did not provide you a changelog / cdc

So you have to

(a) repeatedly query incremental date ranges, with some overlap (b) deal with late arriving rows

(c) deal with occasional bugs where a bug in a sproc caused the timestamp not to update

(d) query each day for all primary keys and remove rows that had been deleted (but only if you have newer data, otherwise you delete old values but don't have the updates)

(e) Sweat about if this goes wrong on a large table, how can you rapidly determine what state is out of sync? (Idea: how do I "hash" this in a way that hints at where the problem is?)

(f) Deduplicate, as mentioned

Martin Kleppmann's "Designing Data-Intensive Applications" book has some detailed discussions of some of the challenges...

My understanding is that Taalas burned a Large Language Model as analog silicon, leading to something that ran (roughly) 10x faster for 10x less power

...at the cost of it being burned in forever. Non-programmable.

Agree, copy-paste feels like it works 80% of the time, which is maddening when 1 in 5 copy paste attempts fail. (mostly from zed to something else) Why??

At least Excel has the courtesy to telling you "Hey the copy-paste buffer said it didn't catch that, can you try again?"

Linux Wayland (Ubuntu 26.04)

> In a compute-constrained world where model demand is growing faster than capacity

I don't buy it.

There have been recent weeks where some of the mid-level models (Hy3, Laguna M.1) are free (true for parts of June and July, see Hy3 in Cyan) . Even then the total token usage appears to be reaching a steady-state.

https://openrouter.ai/rankings#top-models

^ the first graph is tokens per week across all models

I guess we just can only throw ideas at an LLM at a certain rate.

I still have ideas and now I can have an LLM vibe code what I want, but I'm not going to let an agent just run unattended for longer than a few minutes or a few bucks for hobby projects.

So maybe it is a matter of lowering the cost of an LLM so I can let it churn for hours at a cost of pennies... But I suspect demand for tokens is very price-elastic.

Why settle for sampling when you can have the whole dataset?

The spiral pattern is an excellent example. "Sure it looks like this when you zoom out, but when you zoom in, you can see the finer structure of the points..."

Gleam ?

Elixir's exhaustive pattern matching and gradual type checking is also on my radar.

I guess I don't know what you mean by effects typing.

Signatories appear to group into

(a) sellers of hardware (Nvidia, IBM) and those who rent hardware out to others (Amazon, Microsoft)

They want open models because they don't have to license it to run it on their hardware, so they can offer lower cost and inspect what they are offering to clients

(b) open model creators (Arcee, Perplexity, Mistral, IBM, Microsoft)

...who would be unable to work if open-weight models are banned

(c) Software-heavy companies ( Mozilla, The Linux Foundation , ServiceNow, Microsoft, IBM, heck all of them)

...who want models that are cheap to run because that reduces the cost of running an LLM (whether you are trying to replace a software developer or just assist them, the value of a reduced cost LLM is directionally the same)