Comment by pizza234
21 hours ago
The singularity, as defined by Hinton (and others) as RSI (Recursive Self Improvement) may actually be beginning already, as OpenAI has announced an AI acting as a "research intern" (!).
21 hours ago
The singularity, as defined by Hinton (and others) as RSI (Recursive Self Improvement) may actually be beginning already, as OpenAI has announced an AI acting as a "research intern" (!).
How is this different from arguing that Microsoft Clippy was RSI? An AI tool being involved in the process of work can't be the bar for RSI.
I don't think there can be a coherent definition of RSI unless people lay out their theory for how intelligence scales. LLM-assisted coding is great but respectfully optimizing pytorch features or whatever is not gonna lead to exponential improvements. That approach to scaling diminished years ago, leading all the labs to switch to reasoning.
Now it seems reasoning is also yielding diminishing returns, so all the labs are pivoting to specializing in particular fields like math / infosec / biology. They're improving due to accessing new proprietary training data and doing RL with human experts. Again I don't really see any amount of "AI research interns" leading to an exponential improvement to this strategy, they're not the bottleneck in the first place.
>Now it seems reasoning is also yielding diminishing returns
Is the diminishing returns in the room with us?
>so all the labs are pivoting to specializing in particular fields like math / infosec / biology.
They're not pivoting to anything. The goal has always been creating a machine that could automate all or nearly all human work. They're just coming along on that mission.
As for RSI...I think the term is a bit odd in the modern context. It was created at a time when conventional wisdom was that generally intelligent machines would be these logic automatons that could "alter their own code". Instead we have massive neural networks that take months to train.
In this paradigm, the ways a LLM could "improve itself" would be altering its own weights directly or creating and training better, vastly more efficient architectures for the next generation of models.
The former is probably not happening but the latter is possible.
Yes, diminishing returns. Not overall, they've still been able to create more intelligent models even up to today. But the strategy for scaling that intelligence has shifted. From the initial ChatGPT release to GPT-4.1, they were basically scaling up compute training compute / model size. Then 4.5 flopped, while o1 demonstrated that gains could continue by reasoning (scaling up compute at inference time). o1 is now the ancestor of all their flagship models from GPT-5 on.
This is why I'm trying so hard to drill down on the theory of scaling, and not just talk about improvement in general, hand-wavy terms. If the bottleneck of current scaling strategies is training data, or something fundamental about the model architecture, then just throwing more harnessed chatbots at it won't lead to an exponential increase in performance.
Now you could argue that the AI we have now will help us find that change in architecture, and I would agree. But that means we're firmly outside the singularity for the time being, and what people are in fact talking about is a hypothetical.
2 replies →
> Now it seems reasoning is also yielding diminishing returns
Not true. On the contrary, LLMs are developing faster than predicted. They were expected to solve a Millennium Prize by 2030... and here we are in 2026. Release cycles are getting faster. Just compare the most recent GPT or Claude with what they were an year ago.
> How is this different from arguing that Microsoft Clippy was RSI?
We can argue about semantics, but that's not really the point. The point is that what started now - which no doubt is in its infancy - will result in full autonomy quite soon (they project an year or so), with the risk of RSI causing agent development to slip (long term) outside human cognitive control/capacity.
Again, can you lay out your theory for how intelligence scales? You're using a lot of terms like "full autonomy" without definitions. Why do you think that just throwing more harnessed LLMs at (something?) will lead to an increase rate of improvement?
I feel like I laid out several cases where other things were the limiting factor on improvement and more agents wouldn't have helped, and I didn't get a response to those cases.
What "they project" (the labs) is of minor interest to me. Aside from their incentives and track record of lying, in recent months they are laying out a story that is pretty much just the plot of Terminator, and directly referencing rationalist beliefs that were published long before LLMs even existed.
How is it improving, that would require rearranging its weights and biases which it cannot do easily or quickly.
Self improvement during training, and AI self training are already happening. Easily/quickly are seemingly a factor of how much power/hardware you want to use at once.
With the level of compute they have they aren't stuck with frozen models like you are.
The infrastructure provisioning alone to train is heavily dependent on humans, as is dealing with failures (training runs fail a ton). Its not as simple as adding another ec2 on your dashboard. < 1k people in the world know how to do this, there will not be "recursive" or looped continual training for a long long long time. There are so many delicate inputs and controls. Not to mention the chains of businesses and the people required to operate them just to obtain the data needed, clean it and hand it to the llms.
The llms are supervising rlhf and creating synthetic data (to an extent) but they're nowhere close to being able to operate the full training stack end to end. This is a fantasy being sold to investors to create fomo.
Remember they're also limited by an effective memory of like 500k words a turn. Memory systems are lossy, so are swarm/sub agent mechanism. Im not worried about llms becoming self powered super entities anytime soon.
Is easily and quickly a requirement? Isn't it enough that over time it improves itself even if the process is complex and slow?
Do we know it's actually improving itself? Perhaps it's just opaquely sorting all ones and zeros for better lookup efficiency.