@Dorialexander on Backscroll

14 tweets from @Dorialexander that Backscroll ranked for substance.

Préférable de passer par un accès terminal pour ce genre de tâche (cowork/claude code). En version en ligne, trop de trucs bricolés pour la gestion de fichiers : plus un problème d’interface (harness) que de modèle.
0 likes · 0 retweets · 0 replies
Blogpost: https:// pleias.ai/blog/local-ai- for-knowledge … We also open source our library for cache augmented generation on raspberry pi. https:// github.com/Pleias/pi-cach e-augmented-generation …
8 likes · 1 retweets · 1 replies
We describe several practical use cases on the field, including a legal assistant for conflict-related sexual violence (CRSV) survivor networks, powered by a retrained version of Baguettotron (600M variant) on a synthetic environment in humanitarian law.
10 likes · 0 retweets · 3 replies
Ok it’s really synthetic simulation everywhere from now on. Bye rephrasing.
6 likes · 0 retweets · 2 replies
Fable is a good reminder that the frontier of capability is not the frontier of economic viability. Until they can figure out inference margins and alignment trade-off, it’s roughly in the same awkward spot as Opus until mid 2025.
2 likes · 0 retweets · 0 replies
>indefinite compute >likely largest model ever >good arch research (and copying deepseek is free) >come on, do your agent thing >… what did you try to model exactly, anon?
3 likes · 0 retweets · 0 replies
It’s more like an extreme form of genre inertia: straight pulling from romance written in late medieval/renaissance era.
2 likes · 0 retweets · 0 replies
More seriously the main reason it cannot happen and that conditions everything downstream: no tech sector. That’s the critical lifeline for a frontier labs: even after the agentic turn, general economic demand is nowhere near mature enough to cover fixed compute scaling costs.
12 likes · 1 retweets · 2 replies
Everyone is training on synthetic/RL environments now (actually, many drawing from materials coming from Europe). Essentially how pretraining data stopped being a commodity and all about skills.
0 likes · 0 retweets · 0 replies
Yeah vllm/sglang do work quite well know, batching ensures a good theoretical throughout but then everyone want to Claude Code and you have to manage X parallel sessions, each with varying latency.
5 likes · 0 retweets · 2 replies
As much as I’d like to see companies adopting GLM 5.2, I think we underestimate the real unresolved challenge of agentic serving.
8 likes · 0 retweets · 1 replies
as everything, skill issue: you can't distillate efficiently without a good problem design pipeline.
20 likes · 0 retweets · 1 replies
And GLM synth pipelines now plugged on batched dockers (ok singularity since that’s what we can install on cluster — aptly named).
23 likes · 0 retweets · 2 replies
Unfortunately our core issue is much worse than over-regulation: generalized grift. Policy people are daily joining mid companies they helped propel/entrench and it’s totally normalized.
5 likes · 0 retweets · 0 replies

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric — not raw engagement. Curated by Surya Dantuluri.