davedx

I'd expect the memory bandwidth to be the same for the CPU and GPU under a unified memory architecture like Apple silicon uses?

I find it super interesting that we're now in an era where we have LLM's that are quantized to binary weights - 1's and 0's. So effectively they're digital neural networks.

I assume that in addition to the significant memory savings, this should also lead to much simpler matrix multiplication operations? Could models like these run on CPU's efficiently, or does the geometry of the compute mean GPU's are still a better choice?

I did some research on this in the context of self-replicating PV panel construction. I arrived at similar conclusions: mining (ore extraction and refining) was the hardest part. Our current methods involve all involve some kind of high energy system:

- crushing

- breaking down with powerful solutions

- blasting

And a self-replicating probe will (initially at least) be a low energy system. I eventually decided that the pathway with the most likelihood of success would be some kind of very slow crushing/grinding machine that can break down ore into separable components, but then you get into a kind of Darwinist explosive combinatorics research rabbit hole: which crusher/grinder, what kind of machine, how to make something that works on different ore types, what mechanical pressure is better?

Conceptualizing something that can sinter and assemble PV cells was pretty easy, there are broad families of chemistries that work and they mainly differ on input temperatures and output efficiencies. Fairly tractable. But mineral extraction... yeesh, it's extremely difficult.

FWIW on the original article: I think the jump from "insulating wires" to "semiconductor fabs" was kind of obtuse. You don't necessarily need Turing complete PCBs or microchips for most (any?) of this.

This is very critical of an open source project that the maintainer didn't even post here?

"Status:

pgrust is not production-ready yet. It is not performance optimized yet."

The maintainer is not suggesting you use this for anything yourself. So why do you care about spf or (lol) his "discipline in creating the project"?

This YouTube is a pretty good walk through of the gameplay: https://www.youtube.com/watch?v=bxGEMOzWt8w

I agree with this statement. And because it churns less tokens, it's just generally faster too - noticeably, throughout the day, across a range of tasks I get more shit done with Codex.

It's not better at reasoning on complex coding tasks, Claude Opus is still ahead there, but not by a lot.

A bit slower? I think for most of my tasks, Claude takes easily 2x longer for almost everything, even things like just analyzing code. It churns tons of tokens for quite simple things.

IMO that's exactly why it's a bit better at actual problem solving.

You absolutely do not "always have to correct" Codex. I'm not sure what you're doing, but I'd say 80-90% of its edits on my side it doesn't need any revisions.

I switch between both as my daily drivers.

I do almost all my regular coding tasks with Codex 5.5 on medium. Sometimes for niche edge cases, or when I run out of tokens on my Codex sub, I'll switch to Claude. Some recent examples where Claude was able to solve things Codex couldn't:

- 3D gamedev layout: I asked Codex to render a solar system in a certain camera positioning, saying it needed to fit the planets of the system to the viewport. Codex just couldn't do it, even on high reasoning: Claude Opus did it first attempt.

- Tricky Tiptap image drag-n-drop layout implementation: Codex failed this after numerous iterations. Claude Opus also struggled mightily to get it to work, but I think around 3 attempts it nailed it. Both of them ended up grepping the Tiptap code from node_modules - that's the kind of task it was.

But these are really isolated examples. Across all my projects (I have many; mostly TypeScript, but also things like C#), Codex "Just Works" (tm), with minimal prompting effort from me.

It's a bit more nuanced than that. Earlier models definitely benefited a lot more from prompt engineering. I remember this distinctly from building data pipelines to do things like extract data from PDFs over the last year or two - there are numerous "tricks" like negative prompting, including the right number of examples, massaging the mock data in the JSON examples so it wasn't "too realistic", and so on. I saw how this impacted recall by running evals, so it wasn't pseudoscience.

But what has happened is the models have gotten better - which OpenAI is making explicit for some cases in this release. You need that stuff less and less as they become more human and better at inferring what's required implicitly.

You still do need to be explicit, and you probably always will, but you don't need as much "engineering" of the way you're asking for things with more recent models.

Is gemini live the same app I have on my oneplus 15? I don't think that's full duplex?

Some useful context in here: https://attorneygenerallynnfitch.com/wp-content/uploads/2024...

"Meta knew what it was doing"

In December 2015, CEO Zuckerberg listed as one of Meta’s goals for 2016: “Time spent [on the Platorms] increase by 12%” over the following three years. And as of November 2016, Meta’s “overall goal remain[ed] total teen time spent … with some specific efforts (Instagram) taking on tighter focused goals like U.S. teen total time spent.”

Between October 2022 and April 2023, Meta’s own internal metrics show that an average of 208,000 Mississippi young adults used Instagram daily and 345,000 used it monthly. In fact, Meta monitored key metrics for Mississippi, including:

• Ratio of teen daily active users to monthly active users: 0.72

• Increase in monthly active users over a two-month period: 7,894

• By 2020, Meta estimated 100% of MS teens were monthly active users of Instagram

A 97-page internal presentation, “Teen Fundamentals,” in May 2020, described its goal as to “look … to biological factors that are relatively consistent across adolescent development and gain valuable unchanging insights to inform product strategy….”

That presentation conceded, “due to the immature brain they have a much harder time stopping even though they want to – our own product foundation research has shown teens are unhappy with the amount of time they spend on our app.”

One internal communication noted that Meta could “[l]everage teens’ higher tolerance for notifications to push retention and engagement,” while another noted that some users are “overloaded because they are inherently more susceptible to notification dependency.”

As it noted in its 2019 internal presentation, “Teen Mental Health Deep Dive,” “Young people are acutely aware that Instagram can be bad for their mental health, yet are compelled to spend time on the app for fear of missing out on cultural and social trends.”

In another internal presentation, Meta employees express concerns about “content on IG triggering negative emotions among tweens and impacting their mental well-being (and) our ranking algorithms taking into negative spirals & feedback loops that are hard to exit from.”

Meanwhile:

> China’s Ministry of Commerce has led meetings over the past month with major AI companies, including Alibaba, ByteDance, and http://z.ai/, to discuss measures that would restrict overseas access to cutting-edge AI models, including models that have not yet been released.

> The discussions reportedly include not only closed-source models but also open-weight models.

> Future regulations could take the form of a tiered framework based on technological capability. Basic open-source AI models may be managed through a filing system, high-performance models may be subject to security reviews, and the most sensitive frontier models may be banned from public release or restricted to use within China

https://www.reuters.com/world/beijing-is-looking-curbing-ove...

The EU is not "teetering on the edge of a crisis", why do you say that?

Why on earth don't the producers of the game implement this? It sounds trivial to do?

That only happens if you're trading very illiquid stocks.

I'd love to read more about how to use this workflow. What kind of top level instructions does this actually work with? Is there an article out there with some concrete examples of how to do this effectively?

Tautologies 'R us

An expressive and well designed language (elixir) is objectively better than a less well designed language like python. Python probably needs more LoC than elixir for the same task. Python is also untyped by default.

I don't really see the moat for frontier AI labs being "more efficient models" although that could help their margins - I think moats will be built by expanding the horizontal and vertical market expansion - like Anthropic is doing the most at the moment

Wan Streamer is a native-streaming, end-to-end interactive foundation model, designed from the ground up for real-time, low-latency, full-duplex audio-visual interaction. It models language, audio, and video as both input and output within a single Transformer: the sequence is an interleaving of visual, audio, and text input tokens with visual, audio, and text output tokens, coordinated by block-causal attention for incremental streaming.

Define compute deficit?

They've been bringing out open weight models competitive with frontier models. How could they do that if they had a compute deficit?

> The standard way to reduce liability risk is to get liability insurance

Exactly! That's what I do in the Netherlands. It's also common to cover this contractually too - you can negotiate where liability falls for many cases.

Getting a limited liability company for a one-person operation is just overkill.

Yeah I don't really understand the part about not being able to invoice either. IANAL but it's the other way around - you need a VAT number to invoice clients in Germany (and the EU), not outside of it. VAT is exempt for clients outside of the EEA.

Also: I've always used a ZZP structure (one man company - Dutch version) for mine, not a BV (LLC), because there's a thing called Professional Liability Insurance. But maybe it's different in Germany? I can't imagine that doesn't exist there though.

The way it failed to maintain its strategy, or even its build plans, makes me wonder if this is something that could be solved via the attention mechanism itself?

Instead of only using attention to focus on the previous token position, could it also do some kind of higher order "temporal attention" planning where it weighs each previous log (game state + intent) checkpoint when generating outputs?

I don't understand the article.

"I’d say this benchmark answers with a resounding, “Maybe.”

Mythos maybe really is better than the other current models at finding security bugs"

Yet in the results, I don't see Mythos?

It seems like a really well researched article with lots of results for other models, yet the title seems to be clickbait because the results don't contain Mythos, do they?

66k

Ryu vs Chun-Li full super parry win: https://x.com/LostMemeArchive/status/2067292940838535561

> A Japanese esports player and author who competes professionally at fighting video games. He specializes in 2D arcade fighting games, mainly those released by Capcom. Daigo is one of the world's most famous Street Fighter players and is often considered its greatest. His longevity is seen as rare in the world of competitive video games.

There's been a lot of discussion on the Interwebs about whether SpaceX's AI satellites are a) viable in an engineering sense, and b) economical. This project does not attempt to answer b) at all, but does try to have a basic look at a): given the FCC filing of where the satellites will be orbiting, what kind of impact will the cosmic environment there have on the specific compute tasks we expect them to be performing (namely, transformer inference)?

Full disclosure: the project was my idea, and I wrote the basic initial design draft, but Claude did all the rest. I definitely learned some things though. Feedback would be very welcome, in particular ideas about what the "software architecture mitigations" SpaceX might try.

But that sounds like the same workflow as Codex or Claude, except Cursor is only a harness without its own model? (Or do they have their own model?)