laydown dev

brawlhalla nerd
love rl & ml

[laydown dev @ldardev]

DeepSeek API peak hours (weekdays):

🇺🇸 PT: 6–9 PM & 11 PM–3 AM
🇺🇸 ET: 9 PM–12 AM & 2–6 AM
🇬🇧 UK: 2–5 AM & 7–11 AM
🇪🇺 Central Europe: 3–6 AM & 8 AM–12 PM
🇮🇳 India: 6:30–9:30 AM & 11:30 AM–3:30 PM
🇨🇳 China: 9 AM–12 PM & 2–6 PM
🇯🇵 Japan/Korea: 10 AM–1 PM & 3–7 PM
Outside these hours = cheaper off-peak pricing

[laydown dev @ldardev]

LLMs aren't bad at coding. They're bad at asking questions.

Fresh results on the IdeaAMBIG benchmark show a massive gap:

• Spotting an ambiguity on their own: 9.6%
• Fixing it once pointed out: 80.6%

Why? LLMs are trained to complete text, not to handle uncertainty. When a prompt is vague, they don't pause they guess the most statistically likely intent with 100% confidence. This 71% gap is the single biggest reason autonomous coding agents derail, while human in the loop dev setups feel magic

The next big upgrade for agents won't be better code generation, but better self-doubt

[laydown dev @ldardev]

@thsottiaux limit reset ?

[laydown dev @ldardev]

@VladAladin okay this is actually a really cool use of Astra lol going through your own dna data with it is crazy (and in 40min lol)

[laydown dev @ldardev]

@_KevinTang Old games getting a second life in VR is honestly one of the coolest things happening rn. Astra got one running on my Quest 3 too and I genuinely wasn’t expecting it to work that well

[laydown dev @ldardev]

DeepSeek is about to do something pretty funny.

V4.1 Flash apparently beats V4 Pro on capability, speed, cost and total task time.

So until V4.1 Pro ships, they're literally routing Pro requests to Flash and charging Flash prices.

“Flash” is slowly becoming a pricing tier, not a capability tier

[laydown dev @ldardev]

“What’s the smartest model?” is becoming the wrong question.

GPT-6 Astra and Claude Fable 5.1 are tied at 53 on the new Artificial Analysis Intelligence Index.

Yet Astra wins agent automation + terminal work, while Claude wins GDPval, SciCode and HLE.

Same score. Very different brains.

[laydown dev @ldardev]

Trying out @spawn right now and I’m genuinely impressed

You can describe a game build it live in the browser and publish it

Creating is currently free, and games are multiplayer from the start

Really promising !

[laydown dev @ldardev]

@harisahmad59 WTF BRO REALLY WORK LOVE YOU

[laydown dev @ldardev]

Almost everything is now 1:1 in the brawlhalla sim

Starting work on a small accurate visualizer

[laydown dev @ldardev]

im making a brawlhalla sim to train an ai who want to see ?

[laydown dev @ldardev]

Recurrent architectures are back in the AI spotlight.

Sapient Intelligence says it has been working on recurrent reasoning since 2025, with HRM and now HRM-Text, an open-source ~1B parameter model.

The wild part: they claim it was trained from scratch on 40B unique tokens for $1,000 in GPU compute.

They’re now building PRAXIST to help researchers experiment with new AI architectures.

Recurrence might be having a comeback

[laydown dev @ldardev]

I just added flip reset reward to my rocket league ai

Lets see what happens

Currently at 81B steps

[laydown dev @ldardev]

I’m building a Rocket League AI

It’s currently at 80B steps and watching it improve is honestly so fun.

Would you guys want to see a video of how it plays? 👀

[DIMENSION @DIMENSION_YT]

Les USA : *investissent 500 milliards sur l'IA*

La France :