WhitneyLand

“That is exactly how web standards work…”

Says who? You keep mentioning this 0.1% threshold yet…

1. I can’t find any reference to that do you have examples / citations?

2. On the contrary here’s a paper that proposes a 3x higher heuristic: https://arianamirian.com/docs/icse2019_deprecation.pdf

3. It seems there are plenty of examples of features being removed above that threshold NPAPI/SPDY/WebSQL/etc.

4. Resources are finite. It’s not a simple matter of who would be impacted. It’s also opportunity cost and people who could be helped as resources are applied to other efforts.

I was expecting to read pearls of wisdom gleaned from all the hard work done on the project, but I’m not really getting how this relates to ffmpeg.

The few chapters I saw seemed to be pretty generic intro to assembly language type stuff.

But that’s not what the review says:

“It’s actually worse at writing than GPT-4.5, and I think even 4o”

So the review is not consistent with the PR, hence the commenter expressing preference for outside sources.

Agents muddy the waters.

Claude Code gets the most out of Anthropic’s models, that’s why people love it.

Conversely, Gemini CLI makes Gemini Pro 2.5 less capable than the model itself actual is.

It’s such a stark difference I’ve given up using Gemini CLI even with it being free, but still use it for situations amenable to a prompt interface on a regular basis. It’s a very strong model.

If you structure your prompt around small, stateless functions with explicit inputs/output, answers may look cleaner, but that advantage is not at the language level. You can do the same in TypeScript et al and get the same results.

I don’t doubt that anecdotally you’ve had favorable results. However, I’m not aware of any research that backs up this claim. Or forget research - I’m not aware of any simple examples / prompts that would show it in a reproducible way across different LLMs.

Corpus size tends to dominate performance.

I’m skeptical of this claim, where is he getting it from?

“Elixir’s foundations in functional programming…make it easier for large language models to reason about, generate, and test code. LLMs struggle with imperative codebases full of side effects, hidden state, and indirection…”

It seems a larger factor in LLM performance on a given language is the amount of code available to train on.

He’s repurposing asynchrony that’s different from the way most literature and many developers use it, and that shift is doing rhetorical work to justify a particular Zig API split.

No thanks.

Mostly SOTA performance at the 3B level. A notable addition to the small but truly open club of models that provide full disclosure, code, recipes to reproduce their work.

Looks like ballpark a million dollars of GPU time if you want to train up one for yourself (4000 gpus/24 days).

Very nice write up that’s generous in sharing their learnings.

This is a solid and positive contribution.

“Reasoning models like o3 might be better equipped to come up with a plan, so a natural step would be to try switching to those, away from Claude Desktop…”

But…Claude Desktop does have a reasoning mode for both Sonnet and Opus.

What’s with Copilot “agent mode” anyway, how does it compare to using Claude Code or Gemini CLI?

“This paper presents a theoretical proof that AGI systems will structurally collapse under certain semantic conditions…”

No it doesn’t.

Shannon entropy measures statistical uncertainty in data. It says nothing about whether an agent can invent new conceptual frames. Equating “frame changes” with rising entropy is a metaphor, not a theorem, so it doesn’t even make sense as a mathematical proof.

This is philosophical musing at best.

On a related note what’s the hesitation for Anthropic/OpenAI to open up their mobile apps to diy mcp servers?

One thing not really advertised about Claude Code is it makes all command line tools plain text/plain English.

“Get data from the last log entry in <file> and post it to <url>”

It looks like the UX freaks out when you try to do something like pinch zoom.

If you don’t intend to support two finger gestures, you should probably at least hook in and disable them so they don’t disrupt the experience.

If someone at Apple said I want communicate in a natural way on video and not really go into TED talk mode would they get in trouble?

Yes. Here’s a great ELI5 intro:

https://youtu.be/_cr46G2K5Fo?si=Q6iEm3m-Nyge3FUW

If we zoom out a bit to one point he’s trying to make there, while LoRA is fine tuning I think it’s fair to call it a more modular approach than base SFT.

That said, I find the article as a whole off-putting. It doesn’t strengthen one’s claims to call things stupid or a total waste of time. It deals in absolutes, and rants in a way that misleads and foregoes nuance.

So, we currently have o4-mini and o4-mini-high, which represent medium and high usage of “thinking” or use of reasoning tokens.

This announcement adds o3-pro, which pairs with o3 in the same way the o4 models go together.

It should be called o3-high, but to align with the $200 pro membership it’s called pro instead.

That said o3 is already an incredibly powerful model. I prefer it over the new Anthropic 4 models and Gemini 2.5. It’s raw power seems similar to those others, but it’s so good at inline tool use it usually comes out ahead overall.

Any non-trivial code generation/editing should be using an advanced reasoning model, or else you’re losing time fixing more glitches or missing out on better quality solutions.

Of course the caveat is cost, but there’s value on the frontier.

I meant more the subject itself rather than this particular textbook, but I’m curious about your opinion in general.

I came to this opinion after taking it in college and not recalling very much in the way of needed prerequisites, but maybe this is a selective memory…

What are some of the biggest things needed beyond algebra?

Think this would be a great course for high school or even middle school. No plug and chug that makes it a grind, plus a great intro to proofs and deeper mathematical thinking.

Is it possible to be an engineer and not think this is cool?

Paper suggests these changes might actually represent *worse* sleep quality. It says "increased brain entropy during sleep has been linked to hypertension and early-stage Alzheimer's disease" and question whether this represents "a deterioration of sleep quality."

No?

Im not sure what you’re disagreeing with. It seems like selective quoting to sidestep my point about quality of predictions and pivot to ridiculing the original theory which I actually was originally critical of.

I don’t feel any need to shame the author. There’s clearly a good faith effort to contribute and I think the most constructive feedback is to suggest how it could be stronger.

To be fair, predictions are the final gold standard, but not all predictions are equal.

The math would allow for the predictions to be precise, quantifiable, and directly falsifiable.

As is the predictions qualify as interesting, but there are also weaknesses. Some of it was already predicted by others, some needs more verification, some of the claims were more broad “lots of jets, lots of quasars” so they say less than more precise predictions would.

There’s no math.

Modern cosmology requires simulations, simulations require mathematical models.

It’s well researched and points out legitimate shortcomings in current theories.

But without the math you don’t know if everything is really adding up and we’re kind of left with cool story bro.

Why all the trouble to speed things up while at the same time using bfloat16?

I generally agree with the article and the approach given practical constraints, however it’s all stop gap anyway.

Using Gemini 2.5’s 1MM token context window to work with large systems of code at once immediately feels far superior to any other approach. It allows using an LLM for things that are not possible otherwise.

Of course it’s damn expensive and so hard to do in a high quality way it’s rare luxury, for now…

The article reads fine to me.

Yes by technicality RAG could mean any retrieval, but in practice when people use the term it’s almost always referring to some sort of vector embedding and similarity searching.

That’s not a “style” of argument, it’s basic logic. To the extent you can’t define something the strength of your claim that it does not exist is correspondingly weakened.

It’s promising but they need to find a way to add reasoning techniques to DLLMs.

Were a ways out still from matching from frontier performance.