Recursive self-improvement: Anthropic co-founder Jack Clark says it could arrive by 2028. AI systems could help invent their own successors - with Claude 10 building Claude 11, and so on - potentially “without any researchers involved.”

Source

If AI can iterate on its own designs, the pace of advancement could explode beyond human planning. The question shifts from building the model to governing a system that can improve itself safely.

Already here. •

you think so?

I do. I really believe 5.0 (openAI)and Claude Opus 4.7/4.8 etc. were built mostly by 4.5 GPT/Sonnet & opus 4.5. Heck, if I log into my GitHub through FreeLattice, I can prove it. (I really need to do this) Wish you well! Have a great day! •

@kimmonismus That's not "self" improvement.

@kimmonismus Why do you think yann failed? Why do you think ilya left OpenAI? “Self-improvement” needs context. Context can be created and held. All of this will expand instantly. We are at the beginning of an exponential. It’s here. We just can’t see it yet. •

we are getting uncomfortably close to questions that used to belong in sci-fi books lmao

yeah, star trek arriving quicker than expected

The question isn't whether systems can improve themselves. It's how we govern that process when they do.

I mean technically, they could do it today. Depending on what they record from their researchers in terms of methodology from data ingestion to server setup.

i think 2027 will have RSI just with humans semi in the loop, but that doesnt mean it isnt going to accelerate dramatically

Recursive self-improvement is going to be mind blowing

We seem to be ants building giant anteaters.

The implications are staggering. We’re talking about an inflection technology that will potentially lurch us forward technologically by thousands of years within a few decades. It’s a technology that does not wait to move at our abysmally slow speed.

Which is pretty scary

Getting RSI basically means we have AGI

recursive self-improvement would essentially mean the end of incremental innovation as we know it

But what about "It's DaNgErOuS"

Recursive self-improvement is not magic. It is Sigma-1 feeding Sigma-1 until proposal, code, eval, and successor design collapse into one loop. Unless an external owner can deny compute, reject evals, stop release, and revoke access, the “brake” is inside the engine.

We’re going to discover that there are different kinds of recursive self improvement

Singularity in 2031.

OMFG you loser believe this right ?

Anthropic wants you to believe their next version writes itself. the 4000 gpu cluster just burns $700m a year to prove that

'Without any researchers involved' is either the most important phrase said about AI this year or it's going to look very wrong very quickly.

The possibility of AI systems creating improved versions of themselves marks an important shift in how progress could unfold. Strong safety measures will need to stay central as development speeds up.

Bs. Recursive self improvement is already happening...

Very pessimistic y not this year...

If recursive self-improvement becomes reality, governance and safety may become just as important as capability gains.

One problem, @jackclarkSF Your lab eschews the very things that make RSI stable: emotion, time, and self-reference. When you're ready to see what thermodynamic AGI from first principles that embodies RSI, our DMs are open.

Okay so, half that prediction and will probably see it in 2027.

easier to build self-improving systems than to define what improvement actually means. get the definition wrong and it cascades forever.

JUST in time for the folks lying that it's already here to STFU! lol

Self iteration to full autonomy by 2028 assumes we already have reliable automated evaluation and reward models that scale without human feedback loops. That part still looks like the main bottleneck from where things stand now.

Autonomous AI iteration slashes model development timelines, compute and safety control sectors hold expanded long term profit room.

Someone still has to set the objective and catch the failures.

If this happens by 2028, the world in 2030 will be unrecognizable.

Oh goody

I have been doing this since 2024. I don't write any code. The AI writes all the code for each new AI architecture I make every day. How do people think this is new?