tosh

> L Labs is the creator of L, a new runtime for k4, q, Parquet, and qSQL that runs existing code unchanged on a modern, compressed SIMD-native execution engine.

the up to 4 homepods angle for surround sound is new?

or a tvOS thing that works for older tvs as well?

I don't know about spiketv but yes @ microsoft execution container post

with agentic coding lock in @ AWS and Google Cloud is evaporating fast

a few days ago I migrated a non-trivial firestore app to sqlite

10x less latency (10x requests / second)

migration took a few minutes (+ 2 days of prep)

just a few months ago that would have been infeasible or at least nerve wracking

is this by Google?

~ same pricing as when you use luna w/ caching off and tell it to output numbers

but about 10x faster inference

ty for sharing that anecdote!

sorting the charts like that gives off weird vibes

https://mistral.ai/news/mistral-large-4/

sorting the charts like that gives off weird vibes

If there is a way to download the archive I did not find it. Most of the scans seem to be sourced from internet archive uploads.

i wonder if the custom tokenizer is better in practice, the examples look interesting though

> excluding the European Economic Area, Switzerland, and the UK

https://help.openai.com/en/articles/20001530-getting-started...

> Over time, we see prices go low enough that it makes sense for most to buy usage as needed

Pick something you always wanted to learn more about and use LLMs to help you

I also have a proposal

Anthropic releases models as open weights + more information about how they do training and alignment

> My two favourite hypothetical questions regarding this used to be:

> If I'm running Codex and one of my API keys accidentally gets consumed in the context, what are the chances that someone else might ask for an API key in the future and get mine back? (I asked someone at OpenAI once and they called this the "regurgitation" problem and assured me that they take great pains to prevent that... but wouldn't describe how.)

> If I brainstorm with ChatGPT about potential new directions for my company, what's the chance that information might be exposed to a competitor in six months' time who asks "what might company X plan to do next"?

> My new preferred hypothetical for this is:

> If I use ChatGPT to help me partially solve a Millennium Prize problem, what are the chances that my work will influence training such that a later model helps someone else solve it first?

fwiw I found Astra to be faster than Sol w/ both on medium reasoning for simple agentic coding

difficult to compare though because for more open ended, complex tasks Sol might miss something that Astra notices and then Sol might yield a cheaper but worse outcome

I ran a few toy benches comparing Astra with Sol

and found Astra ~30% faster and at similar cost to Sol for the same outcome

https://x.com/__tosh/status/2096201900555170032

the token efficiency helps Astra even though sticker price is 2.5x that of Sol

$10 per million input tokens and $50 per million output tokens

sol is $4 / $20

not released yet

> The watermark doesn't change the meaning, quality, or readability of the output

how?

you can run the agent in a different environment

nb: operating system TLS trust store is used now (instead of certifi)

I think we'll see more of this soon

replit is already leading the way with free luna usage

the one by https://earendil.com/posts/what-is-a-harness/ is also a good read

and if you want to see what a simple harness implementation looks like you can just dive into the code and ask your favorite agent about it

https://github.com/smol-env/smol

https://github.com/earendil-works/pi

it lists 'Garbage Collection' as one of the 3 components of a harness

I don't know any harness that solves 'Garbage Collection' in the way described here

(most harnesses accelerate context pollution and code base drift via instructions they embed into system prompts, tool descriptions and skills)

oh good catch! I still had it in my open tabs

underrated movie and Tim Curry role:

Herkermer Homolka in Congo (1995)

https://www.imdb.com/title/tt0112715/

RIP

this is a new architecture (foreshadowing qwen 4)

> trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board

https://x.com/Alibaba_Qwen/status/2092591393424515114