Been testing local models (Qwen3.6 + opencode).
The 27B dense still edges out 35B MoE on quality
but the speed gap is wild. I get ~7x more tokens/sec output with MoE.
At that point, dense models for local inference start to feel hard to justify.
#LLM#LocalAI#Qwen#MoE
@pchapuis@Fontaine_B@PoitrinalG@bertrandmoine@pbeyssac@atlantico_fr Thnks for sharing! Le chiffre de 80 % est sorti du chapeau à mon avis. Dans le tertiaire, on diminue/arrête bien le chauffage la nuit, mais on le redémarre beaucoup trop tôt en moyenne pour être certain d'atteindre la température à l'arrivée des occupants.
These days I'm having Codex rewrite some of my old Python scripts in GO, just out of laziness of dealing with the virtual env and the dependencies they had 😇
#Gemini CLI struggles to plan; it just wants to code... even in planning mode! Makes agentic coding pretty frustrating. Had the issue 3.0, still there in 3.1.
A shame given how good are the generated artefacts on the gemini's website you would think it's a great coder too.
@eevblog Mate, if someone edited your video, they’d cut the part where you repeat yourself 😊 Don’t take this the wrong way 🙏 Learned a ton from your content 10 years back, and it might be through repetition 🤷
#Moltbook hype was wild this weekend. Here's what quickly became clear:
- Crypto tokens being promoted
- Humans using LLMs to cosplay as "Clawdbots"
Self-fulfilling?
We now have thousands of LLM agents running with root access, exposed to the internet : #promptinjectionFest
OK, #moltbook might just be a clever stunt to launch the $MOLT token.
But the idea of LLM agents with root access across thousands of interconnected machines made me realize how fast things could get weird.
Backing up my data and taking storage offline #OfflineBackup
😤💀ABSOLUTELY NEVER EVER, IN NO CIRCUMSTANCE" ...agentic coding sometimes feels like leading a group of excited 8-year-olds to a baseball game #claudecode#llm
The timeline of software development:
Unit tests → Integration tests → Continuous Deployment (run all tests automatically) → Test Driven Development (write tests BEFORE code)
...and then we invented LLMs so we don't have to bother with any of that anymore 💀
Funny revisiting old #JupyterNotebooks, what once helped me think, now feels obsolete. With coding agents, why record my process when I can just generate the finished script and detailed docs?
No breadcrumbs needed 🤷
I'm now spending my time trimming docs, for years I was begging for them 😀 #Claude is the first developer I've ever had to tell: "too verbose documentation"
With AI, software development costs are trending toward zero. Value shifts to accountability.
Sure, you can build my B2B app; but so can #claudecode. Can you guarantee compensation for revenue loss from bugs? Who's ready to take responsibility?