(Last updated September 13th, 2026)

Enough folks have asked for my POV so here it is:

Yes, be worried:

  1. We have not seen actual evidence of RSI, and its unclear how we would be able to quantify that development past vibes. The public would need to see fully autonomous R&D. The labs currently say they have line of sight, but aren’t there yet.
  2. Models are getting better, faster and are helping create newer models, accelerating development timelines.
  3. They are moving faster than most well-informed people expected, and the cost of intelligence is continuing to collapse.

The economics are probably good (and will get better):

  1. Its well established that model serving can already support decent gross margins even given highly variable compute costs, and developing them them is likely profitable already on a per model basis (fully loaded).
  2. Inference cost per unit of useful output (serving the models) will continue to falling through better hardware, utilization, and standard hardware and software engineering optimization (it is still early)
  3. Cost to train new models will continue to drop drastically (time to train is already going way down, distilling showing high value, new architectures and methods will continue to develop) outside of people costs.
  4. R&D costs will also continue to fall as partially automated R&D, generation of synthetic data, and automated evaluation move to LLMs.

It’s not marketing:

  1. Imagining there is mass collusion between all labs, high level employees etc to turn this into a marketing stunt to drive up IPO prices is assuming a level of multi-year cooperation that would be insane to assume even within one company. 
  2. Trying to frame this up regulatory capture misses the point entirely. Pacing the frontier at the margin leaves room for competitors, open-source, new co’s to start up.
  3. Calling for regulation, and committing to slowing your model development (and therefore assumingely, revenue growth) pre-listing is best modeled as short-term costly. Assuming race dynamics, this becomes long-term costly as well.
  4. Anyone suggesting the companies self-regulate is asking them to, either 1) collude (against the law) or 2) asking them to purposefully behave irrationally (regardless of whether you believe their goals are purely financially motivated or not), and neither solve what is EOD a market dynamics problem.

Open-source makes the this more real, risky, and inevitable:

  1. At this point its clear that the open source labs are both innovating in their ability to find novel (cheap) ways to do pre-training, which sometimes includes distilling frontier lab models (or serving them directly as a wrapper).
  2. Open-source American and Chinese models are capable and useful today, and are pushing pricing towards inference costs, with the gap in capabilities between the two shrinking over time.
  3. Open-source models on local hardware is effectively un-regulatable. The users working with, and the cloud providers who serve them will be (more on that below).

AI Safety views are correct, if alarmist:

  1. This “alarmism” is not in any way new for the labs - individuals now involved with the labs have been saying this was going to be needed for years. Their incentives have changed (now directly benefit from LESS safety-ism), and tune hasn’t changed.  
  2. On the flip side, we should expect people closest to the technology to have the most “psychosis”, expertise at fringe breakthrough science seems to generally lead to irrationalism about its downstream or potential effects.
  3. We should also expect people who had prior beliefs that this technology was going to be transformative (and dangerous) to be the ones working at the labs in the first place 

Self-regulation is a bad outcome, but a good and costly start:

  1. It’s fair to observe that “calling for regulation” is cheap, and has likely more PR value on net. Proposing mechanics, committing to them, and then actually self-regulating is costly.
  2. Compared to industries we regulate today, we are doing nothing for ... possibly the most consequential technology of all time? Companies trying to propose self-regulation is not new. Companies (correctly) looking at global market dynamics to push back on regulation is not new.
  3. The right model for regulation is probably closer to banks alongside a group of independent watch dogs, not one or the other.
  4. We have precedent here, but we have not figured out how to hold agents, companies, or inference providers liable for the actions of their agents. Right now that seems to be mostly a lack of enforcement, but we should expect to need new paradigms regardless.

Why now:

  1. If anything, we have seen recent and speedy convergence of politics and public opinion towards AI safety views acknowledging the danger + lack of potential benefit (in my view anti-datacenter views are a direct manifestation of this)
  2. A series of hacking incidents + advances in ability to solve math problems well ahead of predicted timelines, and political interest have pushed this beyond containment. 
  3. We have an election cycle coming up aligned with planned IPOs of these companies, at valuations that are outrageous and this will be the main story in this election (and the next).

I think the most reasonable assessment of all this is 1) things are getting weirder faster than anyone thought 2) companies are proposing to do something that many have suggested would be necessary for a long time and 3) the timing is primarily a reaction to public sentiment and electoral considerations.  

If you are looking for a conspiracy to explain the last week, my favorite is here.