What leads you to believe that?
> AI labs may also be able to reap tremendous benefit from these inference-scaled models by using them as part of the training process. If so, the large scale-up of compute resources could go into post-training rather than deployment. This would have very different implications for AI governance.
> ...
> So iterated distillation and amplification provides a plausible pathway for scaling inference-during-training to rapidly create much more powerful AI systems. Arguably this would constitute a form of ‘recursive self-improvement’ where AI systems are applied to the task of improving their own capabilities, leading to a rapid escalation.
So "inference scaling is required to scale capabilities" doesn't mean that we're reaching the top of the S-curve in intelligence. If anything, it could mean a shorter timeline and more unpredictable landscape for governance (e.g. due to securing weights no longer as effectively preventing escalation, more in the article).
On its own it wouldn't. But that article came before the later article https://www.tobyord.com/writing/hourly-costs-for-ai-agents which adds the claim that inference (along with everything else being employed at present) is scaling poorly with increasing task lengths. Now maybe the December 2025 claim is wrong, or maybe things will change soon, but the February 2025 article surely doesn't establish either of those.
Even then, you can just compare the progress in open models. Leaps and bounds from where they were 6 months ago.
Now we take for granted that the latest models can juggle between multiple browser tabs, applications, databases, simulators, docker etc to write, execute, e2e test and deploy full-stack applications over hours managing up to dozens of subagents, relatively untouched, without taking down prod even 1% of the time
Not only this, but in the GPT 5.0 era, agents had 0 taste. Nothing looked good. It was the agentic version of the twitter bootstrap era, but worse somehow. Now, I would argue the average agent frontend beats the average human frontend. This isn't even getting into 3D applications in the GPT 5 era
Anyway, the models now reliably execute more than a human can fit into their own context. It's magic
Once we have something that experiences a desktop interface more like a human does, an entire swathe of tooling that has heretofore been nigh-impossible to automate moves into the fold, and that'll be another explosion of folks finally getting to join the agentic workflow world on their industry specific apps...
Literally every major company that has embraced AI coding has suffered devastating downtime this year as a direct result of AI induced failures.
The only companies sticking with AI at this point are the major players who have chasing their masssive overinvestments to the bitter end, crappy coding shops, and for some reason Starbucks and they'll all suffering badly in a year their customers begin demanding massive amounts of human coding to repair all the issues with the AI code.
The popular thing is now to setup loops (eg I setup hourly integrations for Claude/Codex to 1) scrape my Linear, claim achievable tasks, and push PRs or 2) do root cause analysis on customer issues that evaded automated filters, to name a few)
Though for me, my setup still feels mundane. I have AGENTS.md, CLAUDE.md etc and a few skill files. These are purposefully light - tons of examples online you can pull from online. Mine are fairly personal to my setup and products.
Importantly, I also allow Claude and Codex to bypass permissions. Yes, there is a risk they wipe my machine. The productivity upside has been worth it, for me (haven't been burned yet, ~9+ months into running models this way, I have backups, use cloud etc).
As far as maintaining quality, one of the most helpful guardrails over the past year, for me, has been requiring my agents to pipe their changes to local reviewers through OpenCode, Cursor, etc agents to have a council of models with different biases reviewing the changes, and autonomously working towards a completed objective. No matter how good Claude or Codex gets, for example, I will probably always want a different model checking its work. Like GLM, (now with 4.5) Grok, Composer.
Several OpenAI, Anthropic, xAI employees, and popular AI engineers post on X and share helpful tips & updates. Highly recommend for keeping a pulse on startups and AI. I haven't found something close, honestly, other than when I spend time in SF talking to people.