> daily_signal(2026_11_09)
The author of AI's "Bitter Lesson" says raw scale won't get there, Karpathy shipped a site in a weekend with Claude, Elvis's and Eminem's camps sued Suno, and agents made science code portable.
PickBits Daily Signal · Monday, November 9, 2026
// tl;dr
- The researcher whose 2019 "Bitter Lesson" the scaling camp quotes says they misread him. Richard Sutton argues raw compute alone won't produce intelligence: his lesson favors general methods that keep improving with compute, not a mandate to stop inventing methods and just add GPUs, and he calls synthetic data a "big mistake." Ex-OpenAI lead Jerry Tworek says the misreading "reward-hacked the brains of a whole generation of scientists," while Sam Altman and Dario Amodei, both with a financial stake, insist scaling is far from done. No benchmark settles it.
- Andrej Karpathy built and shipped a working video site in a single weekend with Claude. Claude Opus 5.5 filtered 43 million Hacker News comments from 20 years into 6,454 community-recommended videos, and by his own account handled nearly everything via Claude Code while he did only the human-gated steps (a cloud key, a server, a domain). He claims it beats YouTube's recommendations, his own unverified read. Impressive from a top engineer, but no guide to what your team ships next week.
- Elvis Presley's rights holders and an Eminem publisher filed a new pair of lawsuits against Suno. They allege the AI music app illegally generates songs and lyrics that mimic those specific artists, a claim about Suno's output, not only about what it trained on, the part of the fight that reaches what users actually make. It's the latest in a run of suits Suno faces, and the complaints are newly filed, untested and unproven.
- Researchers built an AI agent crew that rewrites a scientist's code to run on a different supercomputer. In an arXiv paper, three cooperating agents translate parallel code among CUDA, OpenMP and SYCL, with a fourth step that routes each failure to the right fix; on 30 HeCBench kernels it beat a state-of-the-art baseline's verified-correctness rate in all 18 model-and-direction combinations while using up to 52.4% fewer tokens. It's an early proof of concept on open models, with no runtime-speedup numbers and no production deployment.
Back in June, Sam Altman stood at Stanford and said a whole generation of researchers had held AI back by doubting what scaling could do. This week the person they all quote to settle that argument, Richard Sutton, said they had read his "Bitter Lesson" backwards, and a researcher who left OpenAI put it harder: the misreading "reward-hacked the brains of a whole generation of scientists." It's the same fight we've been tracking since Sutton called synthetic data a "big mistake" back in August, and this week it boiled over, with the CEOs who sell the compute pushing back in real time.
The rest of the day is quieter, and I'd argue more honest. Karpathy built a working site in a weekend with Claude and said straight out the model did most of the typing. Elvis's and Eminem's camps went after what Suno produces, not just what it trained on. And a crew of agents did the boring, useful thing nobody fights about, moving a lab's code off the one machine it was stuck on. The scaling-forever claim is the loudest of the four today, and the only one nobody has actually shown.
Richard Sutton wrote the 2019 essay the scaling camp cites most, and this week he told them they had been reading it backwards.
1. The researcher the scaling boom quotes as proof keeps telling them they're reading him wrong.
Richard Sutton, who wrote the 2019 essay the "scale is all you need" camp cites, says his lesson never told anyone to stop inventing methods and just add compute.
If you're deciding how much of your roadmap to bet on "next year's model will just be bigger," the guy everyone quotes to back that bet says they've got him wrong. Richard Sutton, the Turing Award winner whose 2019 essay "The Bitter Lesson" is the one the scale-is-all-you-need camp quotes, argued in a widely-shared post on X that raw compute alone will not produce intelligence and that his lesson has been read backwards. His actual point was that general methods which keep improving with compute beat hand-built human knowledge over time, not that you should stop inventing methods and only add GPUs. He calls synthetic data a "big mistake," because any simulation of an infinitely complex world is "microscopic," and wants agents with world models that keep learning from their own experience after training ends.
Plenty of smart people disagree with him. Jerry Tworek, an ex-OpenAI research lead now at a lab called Core Automation, agrees the lesson still holds but says the more compute you pour into training, the more the algorithm matters, and that a misreading of Sutton "has reward-hacked the brains of a whole generation of scientists." Sam Altman reads it the other way: a model just cracked a standing math conjecture, he says, so LLM scaling is far from exhausted, and Dario Amodei calls it "a rainbow with no end." Both run companies whose valuations depend on that being true, so take their confidence with that in mind. No new benchmark or paper decides this one. It's really about whose roadmap you're betting on. This week it got loud.
Why this matters: My own read is that the most honest line in this whole fight is Sutton's, because it is the only one with no product attached to it. The scaling pitch is easy to sell: buy more compute, get more intelligence. That's reason enough to be skeptical when the people making it also sell the compute. Sutton is not saying scale is dead. He is saying it is half the answer, and the other half, methods that keep learning after training, is the part nobody can put on a roadmap yet.
Action this week: Plan around "general methods win with compute," because that part is well supported. Treat "just add GPUs and intelligence falls out" as the open bet it is; no benchmark backs it yet. The teams I've seen get burned on this treated a vendor's roadmap optimism as if it were a research finding, and paid for it. So discount any scaling forecast by who profits from it, Altman and Amodei included. And leave a line in your plan for what Sutton is actually pointing at: agents that keep learning after training. That is where the next real jump likely comes from, not a bigger pile of the same.
the-decoder.com: The inventor of the bitter lesson has a bitter lesson for LLM scaling enthusiasts (November 8, 2026)
the-decoder.com: AI pioneer Sutton calls synthetic data a "big mistake" in the face of an infinitely complex world (August 2026)
the-decoder.com: Sam Altman says a whole generation of researchers held AI back by underestimating what scaling could do (June 2026)
2. Andrej Karpathy built and shipped a working video site over a single weekend with Claude.
Claude Opus 5.5 did nearly all the keystrokes, from sifting 43 million Hacker News comments to standing up the server, while Karpathy handled only the steps a human still has to.
Andrej Karpathy, an OpenAI co-founder now doing research at Anthropic, built and shipped a working, hosted video site called HackerTube (hackertube.org) over a single weekend, using Claude Opus 5.5 to filter 43 million comments from 20 years of Hacker News down to 6,454 videos the community kept recommending. By his own account, the model handled nearly everything through Claude Code, from the data analysis to setting up the server, while he did only the steps a human still has to do: getting a cloud API key, renting a server, buying the domain. He then followed it with TristanTube, a 47,000-video site he says was "one-shotted with Claude Code in 3 minutes."
The result is real and hosted: what used to be a multi-week project came together over a weekend. But Karpathy is a world-class engineer paired with a frontier model, and his claim that it "beats YouTube's recommendations" is his own unverified read, with no cost figures attached. This is proof of what the best engineers can now pull off, not what next week looks like for your team. You can ship a real, hosted thing in a weekend now. How much of that holds for your own team is something only your own work will tell you, not his weekend.
Why this matters: The model did the data sifting and the server setup, but Karpathy still decided what to build, which videos counted, and whether the thing was good enough to ship, and he still had to go rent the server himself. My read is that what got cheap here is the typing, not the deciding. The risk is reading this as "one engineer now equals a team" and planning your headcount around a tweet.
Action this week: Pick one stalled side project and try the same split he did, letting the model take the plumbing while you keep the judgment calls and the human-gated steps. The point is to calibrate your own ceiling against a real task rather than against a famous person's weekend, because the gap between his result and yours is mostly skill and taste, which the model does not hand you. I'd be blunt with anyone using this to justify cutting a team: one weekend from someone as good as Karpathy is not a staffing model. Use it as one and you'll just ship the wrong thing sooner.
3. Elvis's and Eminem's camps are suing the AI that makes songs on command.
The new suits target what Suno's AI produces, songs and lyrics that mimic named artists, not just the copyrighted music it trained on, which is the question that decides whether "make me an Elvis song" is legal.
We've watched the AI-music copyright fight build all year. This week it hit the question that actually affects anyone making AI songs. On November 8, rights holders of Elvis Presley's music and the publisher of some of Eminem's biggest records filed a new pair of lawsuits against the AI music company Suno, alleging it illegally generates songs and lyrics that mimic those specific artists. The earlier fights were mostly about whether Suno could train on copyrighted songs; these new suits go after the output, whether an app that will "make me an Elvis song" on command is producing something that infringes the real artist.
This is not Suno's first legal problem, and that is the point. It already faces copyright litigation from major record labels, a German court ruled in August that it violated copyrights and rejected its fair-use defense, and a report found it had scraped millions of songs from YouTube Music, Deezer and Genius. The training fight is old news; what is new is this suit going after the song Suno hands back. These complaints are days old and unproven, and no court has ruled on whether "in the style of" counts as copying for generative music.
Why this matters: If you have ever typed "make me a song that sounds like" a named artist into an AI, this lawsuit is about whether that output is legal to keep and post. Nothing is decided yet, and the suits are days old and unproven. But the question they press is the one every AI-music user has been waving away: "in the style of Elvis" may or may not be legal to make and publish, and a court is about to start saying which.
Action this week: Hold off on posting anything you generate in a specific famous artist's style until one of these suits actually lands. If you post AI music, keep it to your own voice and original prompts instead of a specific artist's name, because the output, not just the training, is now what's on trial. To me the plaintiffs picked the stronger ground: "it trained on my song" is a slow, hard case, but "it will hand anyone a convincing fake of my artist on request" is the kind of concrete harm a judge can see. Watch whether any Suno rival quietly adds an artist-name filter before it has to.
theverge.com: Elvis and Eminem rights holders are suing Suno (November 8, 2026)
the-decoder.com: German court rules AI music generator Suno violated copyrights, rejects fair-use defense (August 2026)
404media.co: The AI music tool Suno scraped millions of songs and lyrics from YouTube Music, Deezer and Genius (July 2026)
Listen to the Daily Signal podcast at pickbits.ai/daily-signal.
4. Researchers built an AI agent crew that rewrites a scientist's code to run on a different supercomputer.
Three cooperating agents translate parallel code among CUDA, OpenMP and SYCL, beating a state-of-the-art baseline on 30 test kernels while using up to 52% fewer tokens.
Plenty of research runs on only one brand of supercomputer, because the person who could rewrite the code for another machine left two years ago and nobody since wants to touch it. That's the problem this goes after. In an arXiv paper posted November 6, researchers Jianuo Shen, Pengyi Xu and Bo Zhang built an AI agent-skill system that translates parallel scientific code in both directions among the three big high-performance-computing frameworks, CUDA, OpenMP and SYCL, automating the slow, expert-only work that otherwise locks a lab's code to the hardware it was first written for.
Three agents cooperate, and the trick is that they check each other: one translates, one reads the compiler errors, one confirms the result actually runs. A fourth step sends each failure to the right fix instead of blindly retrying the whole thing, and a separate offline pass catches the "it passed" that fools the compiler but is really broken. On 30 HeCBench kernels across three open-source models, it beat a state-of-the-art baseline's verified-correctness rate in all 18 model-and-direction combinations while using up to 52.4% fewer tokens. None of this is in a real lab yet, though. It's a proof of concept on a benchmark with open models, and the paper reports correctness and token savings but no runtime-speedup numbers.
Why this matters: We've spent this year watching AI coding agents get good at research software, and they keep hitting the same wall: they can modernize a codebase but can't tell you whether the science still holds. This one actually guards against that: the agents check each other, with an extra offline pass to catch false passes, instead of one model confidently handing you broken code, and that cross-check is the part I'd actually bank on. And if it holds up, freeing a lab's code from whichever chip vendor a grad student happened to pick years ago would save real time nobody signed up to lose.
Action this week: Steal the idea without waiting for the tool: make the system prove its output actually runs before it claims success, which is worth doing anytime a model hands you code or a translation you cannot verify at a glance. Wait for the number the paper does not give before you get excited: a real runtime speedup on a real scientific code, not just a higher pass rate on a benchmark. Every code-translation demo I've tried held up on small kernels and fell apart on a full application, so that is the test I'd hold this one to. I'll believe it when someone runs it on a lab's actual simulation and says what happened.
» What to watch this week
- Hiring behind Sutton's "continual learning" bet. Sutton wants agents that keep learning after training; the thing to watch is whether research money and people move that way or stay pointed at bigger pretraining runs.
- Whether anyone reproduces Karpathy's weekend outside the top tier. His result came from one of the field's best, so the real test is whether a strong but ordinary team gets anywhere near it on a real project.
- The first ruling that separates AI training from AI output. The Suno suits press whether imitative output infringes even while the training question is still contested, and a judge drawing that line would reset the whole AI-music fight.
- A runtime speedup, on a real codebase, for the HPC translation agents. The paper reports correctness and token savings but no speed numbers and no lab deployment, and that is the figure that turns a benchmark win into a usable tool.
Tomorrow's signal lands here.