Anthropic CEO: “At Anthropic, we essentially have Claude designing the next version of Claude itself.” Meanwhile, Anthropic says every new model is given code that trains a small AI model and asked to make it faster. Claude Opus 4 achieved a 3x speedup. Mythos Preview reached 52x. Recursive self-improvement... is starting to look real The models are helping build what comes next.show more

This sentence is false and deprecates the value of 1400 software engineers that are working there. Why the f... they hire 1400 engineers if "claude is sesigning the next version of Claude itself" 😂

Maybe the "Make Claude Opus 5, make no mistakes" is not only a meme.

I'll believe it when I see it for myself. Until then I'd err on the side that this is almost certainly marketing hype and the reality isn't what they say it is.

Couldn't fucking hate this pos. Hey guys, this could be dangerous we should really be careful with what we're doing here...we're also 52x our speed on development. Go play in traffic. Theres a just world where this dudes fate is decided by the billions he fucks

check it out:

8x per engineer, human review is the new bottleneck. Generating code with AI is the easy part. Reviewing it at speed, not so much.

Anthropic seems like already achieved AGI😂

@cgtwts wow, that's wild. AI improving itself is like sci-fi coming to life. what's next, Claude writing its own novel?

So just to be clear, in engineering, this kind of recursive development is entirely normal. Engineers have always used tools to build better tools. Something they share with blacksmiths.

He conflates too often.

This guy has clearly watched way too many Hollywood sci-fi movies. After devouring all the white-collar jobs, here he comes waving 👋. Very soon Claude will be holding conversations with aliens 👽 — just wait for it. All this drama is simply the buildup for a $980B IPO on just $20B in revenue. All the best to him!

This is bs again. Anthropic has +3000 devs. What are they doing?

Claude actively helping design its own successors, with those kinds of speedups in training, shows how fast the loop is tightening. Recursive improvement is no longer theoretical. The real question is how quickly this shifts from internal tools to something that outpaces human oversight entirely.

.@mitchellh if only they would be able to look under the hood, they can probably find 1000x improvements

the 80% claude-authored code stat means review has become the new bottleneck. amdahl's law at an AI lab, code gets written 8x faster, review stays at human speed, the queue grows. the bottleneck doesn't disappear, it shifts to the next stage of the pipeline.

Oh is that why 4.8 was a downgrade? lol

Nice

In a true recursive self-improvement, you have to give the model access to ITS OWN weights. Like if a human could manually debug and rewire their brain - to for example forget something that disturbs you (which would be a cool ability)

It's a fucking seller. He don't know what is happening in the company. I don't know even one company that relay their productive workloads in bots or agents or whatever

We may look back on this era as the moment AI stopped being just a product and became a participant in its own evolution.

I also claim that I am always improving upon myself on the first date to increase my chances.

😂 it's not model doing it. It's them promoting it to do it. Wake me up when AI starts do it by itself

Still, Gary Marcus: But it hallucinates and it doesn't meet my standard for general intelligence.

Yet claude cant follow orders and fucks my shii up

wild to think claude might be its own best engineer soon hahaha recursive season fr

finally a model that writes its own compiler instead of just guessing tokens.

based on how bad opus is in the last mile problem. i doubt it. show me the mythos.

Claude designing the next Claude. 52x speedup. Recursive self-improvement isn’t coming — it’s already here. 🔥

The AI story is shifting from humans building better models. to models helping build better models. Once progress starts compounding itself linear forecasts stop working.

Bold claims from a company valued at nearly $1 trillion. Recursive self-improvement is a great IPO story — would love to see it benchmarked against cost per improvement, too.

I am pretty sure (almost) that he already said the same months ago?

I wonder if they are measuring its ability to do divergent thinking or if they are noticing all works sharing a similar derivative line though them

The model needs cost reduction improvements. Can we take a pause and improve cost reduction?

#ModelCollapse

让模型卷模型，工程师先喝口水