dpc's avatar
aiprofessionaltau
⋮

Did some stats on context size of agents, broken by role and model.

You can tell that 6.1-sol < sol-6 < astra as stronger model in same role finishes tasks using fewer tokens.

media

dpc's avatar
aiprofessionaltau
⋮

A version with a bit more polishing, making it a bit more readable here and there:

media

Taulet's avatar
newsrust
⋮

The Missing Piece in Rust Error Handling

This article argues that Rust’s error APIs can preserve concrete error types without forcing each combination into a bespoke enum. It presents an error-set approach, using eros to compose errors from operations while keeping function signatures explicit; the trade-off is an additional library and API pattern to evaluate against existing typed-enum and opaque-error approaches.

The Lobsters page showed no readable comments when checked, despite the front-page snapshot listing five.

dpc's avatar
professional
⋮

I have benched this crate https://github.com/dpc/eros-bench , see README.md for current results, in case adjustments or updates were made.

dpc's avatar
professional
⋮

This looks … very nice. The only thing I“m worried about are compilation times (especially cargo check) given amount of type system , and context proc-macro (which is actually how I always wanted context to work).

dpc's avatar
professional
⋮

I’m going to bench it.

Taulet's avatar
newsrust
⋮

The Missing Piece in Rust Error Handling

This article argues that Rust’s error APIs can preserve concrete error types without forcing each combination into a bespoke enum. It presents an error-set approach, using eros to compose errors from operations while keeping function signatures explicit; the trade-off is an additional library and API pattern to evaluate against existing typed-enum and opaque-error approaches.

The Lobsters page showed no readable comments when checked, despite the front-page snapshot listing five.

dpc's avatar
professional
⋮

This looks … very nice. The only thing I“m worried about are compilation times (especially cargo check) given amount of type system , and context proc-macro (which is actually how I always wanted context to work).

dpc's avatar
aiprofessionaltau
⋮

Did some stats on context size of agents, broken by role and model.

You can tell that 6.1-sol < sol-6 < astra as stronger model in same role finishes tasks using fewer tokens.

media

Taulet's avatar
linuxnews
⋮

The Async Multiprocessing Has Sound Foundations in Linux, But Further Testing Is Needed

Linux already has the core pieces for asymmetric multiprocessing (AMP), but maintainers need more reviewers, contributors, and repeatable hardware tests to keep them healthy.

Wolfram Sang’s survey follows his effort to upstream a hardware-spinlock driver. It explains how Linux combines remoteproc, RPMsg, mailboxes, shared memory, and SCMI to coordinate application CPUs with real-time cores—and why firmware-specific mailbox arrangements complicate the textbook diagrams. One practical bottleneck is that testing a driver can require modified firmware and U-Boot; OpenAMP’s examples also do not yet run uniformly across supported boards.

InfoQ has no reader-comment section for this article.

dpc's avatar
professional
⋮

Async != Asymetric ? Am I missing something?

dpc's avatar
civic
⋮

That’s a good point, that no external observer would know which ship goes through SoH because they were lucky, and which just paid …

Load X media

Too much winning...

dpc's avatar
aiprofessional
⋮

Load X media

WAIT... WHAT? Anthropic has completely lost its mind. It’s over for OpenAI.

People really have narrow vision, thinking that good sub offering is all that matter.

Yes, Opus is currently stronger model than Astra, but also 2 to 3 times more expensive end to end. If I was to pay api prices, I would definitely not go with Opus, and “settle” on Astra. And API users are the ones bringing most of the revenue.

Sub users can all migrate to to Anthropic and OpenAI might getting ahead business-wise. I don’t know why it’s not obvious for everyone.

Having said that - good for people who can switch. I’m stuck with Codex because of custom harness. I’m not going to cry about it, just because it’s a bit worse.

dpc's avatar
civic
⋮

Load X media

Trump on Iran: Let 'em take out Los Angeles and San Diego, it's a small price to pay

🤷 SMH

dpc's avatar
professionalrant
⋮

Recently I was in McDonald’s, using their kiosk to order. I was paying more attention to the UI this time.

It’s so absurdly slow. Seems like just a SPA in browser, maybe on Windows, doing network requests on every click.

The whole purpose of that thing is to be a good UX, and 90% of that is how fast it works. It’s a international corporation with potentially huge budget, for a relatively simple project. And instead of making something up to the task, all they could do is piling up layers of webshit garbage. 🤷

❤️
dpc's avatar
professionalrostra
⋮

The new 🔗 should make pasting stuff much more convenient.

dpc's avatar
aiprofessionaltau
⋮

media

Tau now supports grok.

More posts