LLMS

→ View original post on X — @aymericroucher

Poolside Releases Laguna XS and M Coder Models on Hugging Face

By

@aymericroucher

–

28 April 2026 18h00

If you've ever wondered what Poolside was up to… They've been cooking!
They just released Laguna XS and M, their first ever public models, on Hugging Face:
– Two coder models: one is 225B params, of which 23B active, the other is 33B-3A.
– hybrid attention: global v. sliding

→ View original post on X — @randal_olson

Claude Detects When Users Switch to Codex LLM

By

@randal_olson

–

28 April 2026 17h49

Someone discovered that Claude knows when you've been cheating on it with Codex. Lucas maintains an open relationship with his LLMs. Source: https://
reddit.com/r/ClaudeAI/com
ments/1sxe46v/claude_knows_when_you_cheat_on_it_with_codex/
…

→ View original post on X — @kimmonismus

User Claims Opus Represents Significant Downgrade From Version 4.6

By

@kimmonismus

–

28 April 2026 17h35

obviously im not being paid to have an opinion on opus. I think its a big downgrade from Opus 4.6.

→ View original post on X — @kimmonismus

Paid Reviews and AI Model Credibility Concerns

By

@kimmonismus

–

28 April 2026 17h34

I'm being paid to say that Opus 4.7 is bad? What are you implying?

Optimize thinking architecture before optimizing prompts

By

–

28 April 2026 17h12

The shift most people haven't made yet: Stop optimizing the prompt. Optimize the thinking architecture before the prompt. A prompt is a tool. A thinking system is reusable infrastructure. This is what "LLMs don't think, you do" means in practice.

Prompt drift in parallel terminals fixed by hand-built markdown system

By

–

28 April 2026 17h11

Translation: prompt quality without persistent context produces drift. A solo operator running parallel Claude Code terminals put it best: "Each terminal has no idea what decisions I made in the other." Their workaround? A markdown file system they built by hand.

Builders admit LLMs can’t think, manually write thinking architecture files

By

–

28 April 2026 17h10

Hundreds of builders just admitted LLMs don't think. Their fix: manually writing thinking architecture files (PROGRESS.md, DECISIONS.md, IDENTITY.md) for every project. Most AI advice still chases better prompts. The real work moved up a layer. Here's what changed and what it

→ View original post on X — @theahmadosman

KV Cache Quantization Beyond FP8 Degrades Model Performance

By

@theahmadosman

–

28 April 2026 16h56

I keep seeing this advice to quantize the KVCache to 4-bit and save on memory Please don’t do that KV Cache quantization beyond FP8 usually is asking for a nerfed and incoherent model

Stanford AI Index: Model race over per Elo ratings

By

–

28 April 2026 16h50

The AI model race is over. Most people won’t realize it for another six months. Stanford’s 2026 AI Index published the numbers two weeks ago. Arena Elo ratings across every major lab: Anthropic 1,503. xAI 1,495. Google 1,494. OpenAI 1,481. Alibaba 1,449. DeepSeek 1,424.