Archo

A person who builds with AI

[Archo @archodeveloper]

@coldniko @unbug Have you looked at Eagle-4?

[Archo @archodeveloper]

It’s because large language models are trained on the average of all ideas. When prompted correctly, especially with Claude models, you can get them to come up with new things, but it’s difficult

The AI he’s describing is something that is not an LLM, but we already have all of the technology to run it

[Archo @archodeveloper]

@RichardSocher @Recursive_SI @coldniko Seems like a great size for Strata 👀

[Archo @archodeveloper]

OpenAI more stores prompt cache for codex *permanently*.

Think about that for a second.

The prompt cache includes prompts, code, etc.

Anything codex sees.

I hope OpenAI means it’ll be kept for 12 hours and they’ll eat the cost afterwards.

We’ll see if Tibo clarifies.

[Tibo @thsottiaux]

Day 16/

You can now cache for 12h in the API and for logged in users across Codex & Work we keep cache forever and you don’t have any penalty for coming back to an old conversation or resuming after a few hours and are now guaranteed to hit the cache.

[Archo @archodeveloper]

@jhanikhil TestFlight please 🙏

[Archo @archodeveloper]

@hunvreus @AmpCode orbs. Zero contest.

[Archo @archodeveloper]

@januarycomputer Being bullish on Mistral is a recession indicator

[Archo @archodeveloper]

@zeddotdev Codemode

[Archo @archodeveloper]

@dummerspast39 Last I checked 135iq is more than 115 unless I’m blind

[Archo @archodeveloper]

@RobertJBye 🙏 I am

[Archo @archodeveloper]

@unbug Try project maya, you should be able to get 50+tps

[Archo @archodeveloper]

@tianyi_gc You planning to maintain this?

[Archo @archodeveloper]

@scaling01 If it’s faster and cheaper per token I don’t really mind the token consumption, it still ends up faster and cheaper overall

Even if the official provider is slow there are others (baseten) with 5x the speed of sol per token, it balances out

And as you mentioned, swarms

[Archo @archodeveloper]

@AdamHoltererer What skills are you testing? Tasteskill v2?

[maria @maria_rcks]

ok but what the fuck is this model name

[Archo @archodeveloper]

@maria_rcks Oh my sweet summer child

[Archo @archodeveloper]

@SSHCodes Time per task is a much better metric than tps because of token efficiency

[Archo @archodeveloper]

@rachpradhan How does it compare to gpui?

[Archo @archodeveloper]

From what I can tell GLM wasn’t as heavily trained for ZCode based on public info, though they do recommend using it there.

On the other hand, in the original DeepSeek V4.1 technical report they stated a harness comparison where DeepSeek harness beat other commonly used harnesses by a considerable amount

In addition, third party tests showed an even more significant jump with the new DeepSeek V4.1 Flash checkpoint

https://t.co/kWT5rgUZ53

I assume DeepSeek Harness encourages more aggressive self verification, but haven’t confirmed

[Paweł Huryn @PawelHuryn]

Turns out DeepSeek harness improves the score of DeepSeek V4.1 Flash by 2-3 points on average.

It also costs 80% more due to subagents and more output tokens.

New results:

DeepSeek V4.1 Flash 1017 + Claude Code harness:
- max, mean of 7: 39.1
- high, mean of 5: 36.4
- low, mean of 5: 34.4

DeepSeek V4.1 Flash 1017 + DeepSeek harness:
- max, mean of 7: 42.7
- high, mean of 5: 39.2
- low, mean of 5: 34.8

For comparison:

GPT-6 Astra, max: 45
Opus 5.5, max: 41.7
Muse Wonder Full 1.0, max: 42

It's a crazy model, and the only one catched cheating. All the results are available on the benchmark website.