paulmist

Openrouter is a terrible proxy for frontier, token leaderboards are dominated by fast models like Deepseek 4.1 or Luna 6.

> Winner right now is Codex by a mile

Opposite in my experience. I need to limit codex to 500k on medium/low, still run out in 2-3 days with 1 CLI window. CC gives me 4-5 medium/high days with 2-3 CLI windows, and Opus is still great for other regular dumb engineering/refactoring.

On the other hand my head starts to hurt if I read Opus for too long, hopefully they fixed it with 5.5.

I think that such fine-tuning hinges on what do you consider to be a mistake, which is context dependent. Having task-specific finetuned models goes against the status quo of generalization/centralization where few large companies serve a limited amount of models efficiently - both due to inference efficiency and the need/want to control the model.

Having a human-like LLM ecosystem with deep specialization requires a paradigm change in how LLMs are trained - and held accountable. How do we put trust in a specific finetuned LLM rather than the institution behind it? Is there any better approach than the very inefficient evolutionary?

Definitely a rough estimate! I recon base rate is also higher amongst military personnel & it’s not uniformly distributed there either.

Assuming a predominantly male population, the suicide rate should be around 20 per 100000 per year. 5 people in a month gives 18x higher rate. Too short period to draw any statistical conclusions but it’s a large blip.

Correct me if I'm wrong, but all of the aforementioned advances were made in the last year? Until very recently few people had access to these tools. Most people still don't know how to use ChatGPT, and very few use tools like CC regularily. If in a few years these frontier tools become commonplace and people upskill we would should see a network effect?

IIRC Taiwan took a page out of Singapore's playbook and went all in on electrical engineering and adjecent fields. It was very much a long-term strategy. Germany probably didn't feel nearly as much pressure, and was already very strong in all industry.

> “Semiconductors, physical AI, and AI data centers are the triple axis for a great leap forward.”

Not the best wording... I wonder how serious this announcement is.

Personal anecdote on ROI - I was at an early stage startup earlier this year where we had some burstable long-running GPU tasks (<100 VMs). Accross GCP and OCI we couldn't get our hands on L40S on-demand, and had to resort to T4s (released 2018). Sometimes even these were unavailable, and we would have a P4 (2016!) fallback. AWS sells A100s (2020) at $4/hr except they don't even have capacity for x1 versions, you have to rent x8.

I think you're confounding streaming with the rise of solo-queue but cooperative (CS:GO, LoL...) games around 2010. Both were enabled by cheaper computing and alllowed competitive ladders to scale to millions of players. Games where you have to queue with a pre-made team (or pro teams for that matter) are considerably less toxic. Toxic streamers are more of a consequence of an already toxic community.

I think that's a normal evolution of these games? In the end they are cooperative so your teammates depend on you. Although you'll find plenty of people at the top of the ladder spamming werid strategies and being successful.