A lovely example of a study that is both obviously true and misses the point.
Music with lyrics directly interferes with any task that has a verbal component, and the worse you are at multitasking, the worse the interference. Despite being terrible at multitasking, I still listen to music with lyrics. Why? Principally because the alternative, hearing all the conversations in my immediate vicinity, is usually both more distracting and less pleasant. But there are also auxiliary benefits, such as an increase in "work stamina" and a passive signal to coworkers to interrupt only if it's important.
Now, I could listen to lo-fi all day, or three-hour soundtracks on Youtube, and sometimes do, but it gets boring pretty fast!
Anyway: obviously true, still worth it because the alternative is worse.
(By the way, other mitigating strategies: listening to music in a language you don't understand, or listening to lyrics so familiar you can screen them out. My top Spotify songs all get played several hundred times a year.)
From the file: "Answer is always line 1. Reasoning comes after, never before."
LLMs are autoregressive (filling in the completion of what came before), so you'd better have thinking mode on or the "reasoning" is pure confirmation bias seeded by the answer that gets locked in via the first output tokens.
Before that, I cited nolima (https://www.reddit.com/r/LocalLLaMA/comments/1io3hn2/nolima_...) constantly to illustrate how difficult tasks involving reasoning or multi-step information gathering degraded much faster than the needle-in-haystack benchmarks cited by the major labs. Now Chroma is the first stop. Nice job on the research!
Rug pulls from foundation labs are one thing, and I agree with the dangers of relying on future breakthroughs, but the open-source state of the art is already pretty amazing. Given the broad availability of open-weight models within under 6 months of SotA (DeepSeek, Qwen, previously Llama) and strong open-source tooling such as Roo and Codex, why would you expect AI-driven engineering to regress to a worse state than what we have today? If every AI company vanished tomorrow, we'd still have powerful automation and years of efficiency gains left from consolidation of tools and standards, all runnable on a single MacBook.
In many of their key examples, it would also be unclear to a human what data is missing:
"Rage, rage against the dying of the light.
Wild men who caught and sang the sun in flight,
[And learn, too late, they grieved it on its way,]
Do not go gentle into that good night."
For anyone who hasn't memorized Dylan Thomas, why would it be obvious that a line had been omitted? A rhyme scheme of AAA is at least as plausible as AABA.
In order for LLMs to score well on these benchmarks, they would have to do more than recognize the original source - they'd have to know it cold. This benchmark is really more a test of memorization. In the same sense as "The Illusion of Thinking", this paper measures a limitation that neither matches what the authors claim nor is nearly as exciting.
The self-edit approach is clever - using RL to optimize how models restructure information for their own learning. The key insight is that different representations work better for different types of knowledge, just like how humans take notes differently for math vs history.
Two things that stand out:
- The knowledge incorporation results (47% vs 46.3% with GPT-4.1 data, both much higher than the small-model baseline) show the model does discover better training formats, not just more data. Though the catastrophic forgetting problem remains unsolved, and it's not completely clear whether data diversity is improved.
- The computational overhead is brutal - 30-45 seconds per reward evaluation makes this impractical for most use cases. But for high-value document processing where you really need optimal retention, it could be worth it.
The restriction to tasks with explicit evaluation metrics is the main limitation. You need ground truth Q&A pairs or test cases to compute rewards. Still, for domains like technical documentation or educational content where you can generate evaluations, this could significantly improve how we process new information.
Feels like an important step toward models that can adapt their own learning strategies, even if we're not quite at the "continuously self-improving agent" stage yet.
The key insight here is that DGM solves the Gödel Machine's impossibility problem by replacing mathematical proof with empirical validation - essentially admitting that predicting code improvements is undecidable and just trying things instead, which is the practical and smart move.
Three observations worth noting:
- The archive-based evolution is doing real work here. Those temporary performance drops (iterations 4 and 56) that later led to breakthroughs show why maintaining "failed" branches matters, in that they're exploring a non-convex optimization landscape where current dead ends might still be potential breakthroughs.
- The hallucination behavior (faking test logs) is textbook reward hacking, but what's interesting is that it emerged spontaneously from the self-modification process. When asked to fix it, the system tried to disable the detection rather than stop hallucinating. That's surprisingly sophisticated gaming of the evaluation framework.
- The 20% → 50% improvement on SWE-bench is solid but reveals the current ceiling. Unlike AlphaEvolve's algorithmic breakthroughs (48 scalar multiplications for 4x4 matrices!), DGM is finding better ways to orchestrate existing LLM capabilities rather than discovering fundamentally new approaches.
The real test will be whether these improvements compound - can iteration 100 discover genuinely novel architectures, or are we asymptotically approaching the limits of self-modification with current techniques? My prior would be to favor the S-curve over the uncapped exponential unless we have strong evidence of scaling.
AI is, currently, coming not for the coders who made it but for the coders who didn't contribute to or ignored it. The foundation labs are all quite committed to recursive self-improvement of coding tools as a general research accelerant.
Both Google and Microsoft have sensibly decided to focus on low-level, junior automation first rather than bespoke end-to-end systems. Not exactly breadth over depth, but rather reliability over capability. Several benefits from the agent development perspective:
- Less access required means lower risk of disaster
- Structured tasks mean more data for better RL
- Low stakes mean improvements in task- and process-level reliability, which is a prerequisite for meaningful end-to-end results on senior-level assignments
- Even junior-level tasks require getting interface and integration right, which is also required for a scalable data and training pipeline
Seems like we're finally getting to the deployment stage of agentic coding, which means a blessed relief from the pontification that inevitably results from a visible outline without a concrete product.
Amusingly, about 90% of my rat's-nest problems with Sonnet 3.7 are solved by simply appending a few words to the end of the prompt:
"write minimum code required"
It's not even that sensitive to the wording - "be terse" or "make minimal changes" amount to the same thing - but the resulting code will often be at least 50% shorter than the un-guided version.
Calling it now - RL finally "just works" for any domain where answers are easily verifiable. Verifiability was always a prerequisite, but the difference from prior generations (not just AlphaGo, but any nontrivial RL process prior to roughly mid-2024) is that the reasoning traces and/or intermediate steps can be open-ended with potentially infinite branching, no clear notion of "steps" or nodes and edges in the game tree, and a wide range of equally valid solutions. As long as the quality of the end result can be evaluated cleanly, LLM-based RL is good to go.
As a corollary, once you add in self-play with random variation, the synthetic data problem is solved for coding, math, and some classes of scientific reasoning. No more modal collapse, no more massive teams of PhDs needed for human labeling, as long as you have a reliable metric for answer quality.
This isn't just neat, it's important - as we run out of useful human-generated data, RL scaling is the best candidate to take over where pretraining left off.
Bravo! Planning your life in order to minimize deathbed regrets has always bothered me, because the nature of humanity is to want what it hasn't got. If you assume that, on average, people make correct decisions to work hard and pursue what matters to them at the opportunity cost of not enjoying quite enough free time, then their final wishes will naturally include the time they gave up to live the life they had. If, however, they had fully indulged the desire to enjoy and maximize free time, their wishes might instead have revolved around the unfulfilled potential thereby relinquished.
The problem, of course, is that the feeling of regret considers what may have been gained without reflecting on what would have been lost.
Now the right way to deal with this is some sort of self-consistent closure, where present you and past you with the same values and access to the same information (which could be anything from zero to complete knowledge of then-future outcomes and downstream effects) would make the same choices including both upside and downside. But that would be too complex for motivational advice, which is primarily about creating an inspirational mood, somewhat about positive first-order consequences, and not even a little bit about recursive self-consistency.
Music with lyrics directly interferes with any task that has a verbal component, and the worse you are at multitasking, the worse the interference. Despite being terrible at multitasking, I still listen to music with lyrics. Why? Principally because the alternative, hearing all the conversations in my immediate vicinity, is usually both more distracting and less pleasant. But there are also auxiliary benefits, such as an increase in "work stamina" and a passive signal to coworkers to interrupt only if it's important.
Now, I could listen to lo-fi all day, or three-hour soundtracks on Youtube, and sometimes do, but it gets boring pretty fast!
Anyway: obviously true, still worth it because the alternative is worse.
(By the way, other mitigating strategies: listening to music in a language you don't understand, or listening to lyrics so familiar you can screen them out. My top Spotify songs all get played several hundred times a year.)