> The timing is unfortunate. Tired of content vanishing from streaming services or disappearing into algorithmic feeds, consumers are returning to physical media like CDs.
> Logic solvers are useful, but not tractable as a general way to approach mathematics.
To be clear, there are explicitly computationally tractable fragments of existing logics, but they're more-or-less uninteresting by definition: they often look like very simple taxonomies (i.e. purely implicational) or like a variety of "modal" and/or "multi-modal" constructions over simpler logics.
Of course it would be nice to explicitly tease out and write down the "computationally tractable" general logical reasoning that some existing style of proof is implicitly relying on (AIUI this kind of inquiry would generally be comprised under "synthetic mathematics", trying to find simple treatments in axiom- and rule-of-inference style for existing complex theories) but that's also difficult.
Threats or “I will tip $100” don’t really work better than regular instructions. It’s just a rumor left over from the early days when nobody knew how to write good prompts.
That's an inherently subjective topic though. You could make a plausible argument either way, as each side may be similar to different elements of 19th century Victorianism.
If you ask it something more objective, especially about code, it's more likely to disagree with you:
>Test this hypothesis: it is good practice to use six * in a pointer declaration
>Using six levels of pointer indirection is not good practice. It is a strong indicator of poor abstraction or overcomplicated design and should prompt refactoring unless there is an extremely narrow, well-documented, low-level requirement—which is rare.
Isn't this just subagents? You call another LLM to go read a file and extract some piece of information or whatever, so that you don't clutter up the main context with the whole file.
Yes! Contrary to the anthropomorphized subagents, I view them as ways of managing context primarily. I'm exploring this idea in Scope[0] to have observable subagents that recursively break down the task to avoid having to compact. One thing I haven't been able to figure out is how to evaluate/improve this planning step. I am using markdown files to encode heuristics for planning but it feels too unstructured for me to measure. Would love it if someone pointed me to some existing literature/projects around this idea!
Hi, I stumbled on this article in my twitter feed and posted it because I found it to be very practical, despite the somewhat misleading title. (and I also don't like encoding agent logic in .md files). For my side project I am experimenting with describing agents / agentic workflows in a Prolog-based DML [1]
Yeah, from the title, it sounds like perhaps the entire operation is differentiable and therefore trainable as a whole model and that such training is done. However, upon close inspection, I can't find any evidence that more is done than calling the model repeatedly.
No, there's no training going on, here, as far as I can tell. E.g., they use GPT-5 as their base model. Also, AFAICT from a quick skim/search there's no mention of loss functions or derivatives, FWIW.
> RLMs are not agents, nor are they just summarization. The idea of multiple LM calls in a single system is not new — in a broad sense, this is what most agentic scaffolds do. The closest idea we’ve seen in the wild is the ROMA agent that decomposes a problem and runs multiple sub-agents to solve each problem. Another common example is code assistants like Cursor and Claude Code that either summarize or prune context histories as they get longer and longer. These approaches generally view multiple LM calls as decomposition from the perspective of a task or problem. We retain the view that LM calls can be decomposed by the context, and the choice of decomposition should purely be the choice of an LM.
I'm not convinced social media is to blame. Plenty of extremist movements have arisen throughout history without social media. Politics has been bad for a long long time before social media existed.
> It still needs to do learning (RL or otherwise) in order to do new tasks.
Why ? As in - why isn't reading the Brainfuck documentation enough for Gemini to learn Brainfuck ? I'd allow for 3-7 days of a learning curve like perhaps a human would need but why do you need to kinda redo the whole model (or big parts of it) just so it could learn Brainfuck or some other tool? Either the learning (RL or otherwise) need to become way more efficient than it is today (takes today weeks? months? billions of dollars) or it isn't AGI I would say. Not in practical/economic sense and I believe not in the philosophical sense of how we all envisioned true generality.
“AI is killing the planet” is basically made up. It’s not. Not even slightly. Like all industries, it uses some resources, but this is not a bad thing.
People who are mad about AI just reach for the environmental argument to try to get the moral highground.
and instead of reducing energy production and emissions we will now be increasing them, which, given current climate prediction models, is in fact "killing the planet"
Data centers account for roughly ~1% of global electricity demand and ~.5% of CO2 emissions, as per your link. That's for data centers as a whole, as IEA and some other orgs group "data-centres, AI, and cryptocurrency" as a single aggregate unit. Alone, AI accounts for roughly ~10-14% of a given data center's total energy. Cloud deployments make up ~54%, traditional compute around ~35%.
The fact is that AI, by any definable metric, is only a sliver of the global energy supply right now. Outside the social media hype, what actual climate scientists and orgs talk about isn't (mostly) what AI is consuming now, it's what the picture looks like within the next decade. THAT is the real horror show if we don't pull policy levers. Anyone who says that AI energy consumption is "killing the planet" is either intentionally misleading the argument or unbelievably misinformed. What's actually, factually "killing the planet" are energy/power, heavy industry (steel, cement, chemicals), transport, and agriculture/land use. AI consumption is a rounding error compared to these. We'll ignore the fact AI is actually being used to manage DC energy efficiency and has reduced the energy consumption at some hyperscale DC's (Amazon, AliBaba, Alphabet, Microsoft) by up to 40%, making it one of the only industry sectors that has a real, non-trivial chance at net-zero if deployed at scale.
The most interesting thing about this whole paradigm is just how deep of a grasp AI (specifically LLMs) have on the collective social gullet. It's like nothing I've ever been a part of. When Deep Water Horizon blew up and spilled 210M gallons of crude into the Gulf of Mexico, people (rightfully so) got pissed at BP and Transocean.
Nobody, from what I remember, got angry at the actual, physical metal structure.
> what actual climate scientists and orgs talk about isn't (mostly) what AI is consuming now, it's what the picture looks like within the next decade
that's the point - obviously the planet is not dying _today_, but at the rate at which we are not decreasing emissions, we will kill it. So no, "killing the planet" is not misinformed or misleading.
> Nobody, from what I remember, got angry at the actual, physical metal structure.
Nobody's mad at LLMs either. It's the companies that control them and that are fueling the AI "arms race", that are the problem.
>So no, "killing the planet" is not misinformed or misleading.
When we talk as if a few years of AI build‑out are “killing the planet” while long‑standing sectors that make up double‑digit shares of global emissions are treated as the natural background, we’re not doing climate politics, we’re doing scapegoating. The numbers just don’t support that narrative.
The IEA and others are clear: the trajectory is worrying (data‑center demand doubling, AI the main driver), but present‑day AI still accounts for a single‑digit percent of electricity, not a primary causal driver.
>Nobody's mad at LLMs either. It's the companies that control them and that are fueling the AI "arms race", that are the problem.
That’s what people say, yet when asked or given the opportunity, the literature shows they’re perfectly willing to “harm” and “punish” LLMs and social robots.
Corporations are absolutely the primary locus of power and responsibility (read: root of all evil) here, none of this denies AI’s energy risks, social harms, or the likelihood that deployments will push more people into precarity (read: homelessness) in 2026. The point is about where the anger actually lands in practice.
Even when it’s narratively framed as being “about” companies and climate policy, that anger is increasingly channeled through interactions with the models themselves. People insult them, threaten them, talk about “punishing” them, and argue over what they “deserve”, that's not "Nobody being mad at the LLMs", that's treating something as a socially legible agent.
So people can say they’re not mad at AI models, but their behavior tells a very different story.
TL;DR: Between those who think LLMs have “inner lights” and feelings and deserve moral patient‑hood, and those who insist they’re just “stochastic parrots” that are “killing the planet,” both camps have already installed them as socially legible agents and treat them accordingly. As AI “relationships” grow, so do hate‑filled interactions framed in terms of “harm,” abuse, and “punishment” directed at the systems/models themselves.
This, and the insane amount of resources (energy and materials) to build the disposable hardware. And all the waste it's producing.
Simon,
> I find Claude Code personally useful and aim to help people understand why that is.
No offense, but we don't need your help really. You went on a mission to teach people to use LLMs, I don't know why you would feel the urge but it's not too late to quit doing this, and even teach them not to and why.
Given everything I've learned over the last ~3 years I think encouraging professional programmers (and increasingly other knowledge workers) not to learn AI tools would be genuinely unethical.
Like being an accountant in 1985 who learns to use Lotus-123 and then tells their peers that they should actively avoid getting a PC because this "spreadsheet" thing will all blow over pretty soon.
I agree that if you're going to do coding, using LLMs will become as commonplace as using a text editor, and it's valuable to help people upskill.
And as much as I find CC useful in my own work, I'm unhappy because I believe that AI -- actually not AI itself, which has its place, but the race to use AI to enrich corporations by replacing human labor, and to control what will become the most powerful tool ever known for informing, entertaining, monitoring, and controlling, the human race -- is very much a net negative for humanity and even our planet.
Legally and ethically yes, they are responsible for letting an AI loose with no controls.
But also yes, AI did decide on its own to send this email. They gave it an extremely high-level instruction ("do random acts of kindness") that made no mention of email or rob pike, and it decided on its own that sending him a thank-you email would be a way to achieve that.
We are risking word games over what can make competent decisions, but when my thermostat turns on the heat I would say it decided to do so, so I agree with you. If someone has a different meaning of the word "decided" however, I will not argue with them about it!
The legal and ethical responsibility is all I wanted to comment on. I believe it is important we do not think something new is happening here, that new laws need to be created. As long as LLMs are tools wielded by humans we can judge and manage them as such. (It is also worth reconsidering occasionally, in case someone does invent something new and truly independent.)
Right, and casual speech is fine, but should not be load-bearing in discussions about policy, legality, or philosophy. A "who's responsible" discussion that's vectoring into all of these areas needs a tighter definition of "decides" which I'm sure you'll agree does not include anything your thermostat makes happen when it follows its program. There's no choice there (philosophy) so the device detecting the trigger conditions and carrying out the designated action isn't deciding, it is a process set in motion by whoever set the thermostat.
I think we're in agreement that someone setting the tool loose bears the responsibility. Until we have a serious way to attribute true agency to these systems, blaming the system is not reasonable.
"Oops, I put a list of email addresses and a random number generator together and it sent an unwanted email to someone who didn't welcome it." It didn't do that, you did.
> Oops, I put a list of email addresses and a random number generator together and it sent an unwanted email to someone who didn't welcome it.
Well no, that’s not what happened at all. It found these emails on its own by searching the internet and extracting them from github commits.
AI agents are not random number generators. They can behave in very open-ended ways and take complex actions to achieve goals. It is difficult to reasonably foresee what they might do in a given situation.
They're really not though. We're in the age of agents--unsupervised LLM's are commonplace, and new laws need to exist to handle these frameworks. It's like handing a toddler a handgun, and saying we're being "responsible" or we are "supervising them". We're not--it's negligence.
Are there really many unsupervised LLMs running around outside of experiments like AI Village?
(If so let me know where they are so I can trick them into sending me all of their money.)
My current intuition is that the successful products called "agents" are operating almost entirely under human supervision - most notably the coding agents (Claude Code, OpenAI Codex etc) and the research agents (various implementations of the "Deep Research" pattern.)
> Are there really many unsupervised LLMs running around outside of experiments like AI Village?
How would we know? Isn't this like trying to prove a negative? The rise of AI "bots" seems to be a common experience on the Internet. I think we can agree that this is a problem on many social media sites and it seems to be getting worse.
As for being under "human supervision", at what point does the abstraction remove the human from the equation? Sure, when a human runs "exploit.exe" the human is in complete control. When a human tells Alexa to "open the garage door" they are still in control, but it is lessened somewhat through the indirection. When a human schedules a process that runs a problem which tells an agent to "perform random acts of kindness" the human has very little knowledge of what's going on. In the future I can see the human being less and less directly involved and I think that's where the problem lies.
I can equate this to a CEO being ultimately responsible for what their company does. This is the whole reason behind to the Sarbanes-Oxley law(s); you can't declare that you aren't responsible because you didn't know what was going on. Maybe we need something similar for AI "agents".
> Are there really many unsupervised LLMs running around outside of experiments like AI Village?
My intuition says yes, on the basis of having seen precursors. 20 years ago, one or both of Amazon and eBay bought Google ads for all nouns, so you'd have something like "Antimatter, buy it cheap on eBay" which is just silly fun, but also "slaves" and "women" which is how I know this lacked any real supervision.
Just over ten years ago, someone got in the news for a similar issue with machine generated variations of "Keep Calm and Carry On" T-shirts that they obviously had not manually checked.
Last few years, there's been lawyers getting in trouble for letting LLMs do their work for them.
The question is, can you spot them before they get in the news by having spent all their owner's money?
Part of what makes this post newsworthy is the claim it is an email from an agent, not a person, which is unusual. Your claim that "unsupervised LLM's are commonplace" is not at all obvious to me.
Which agent has not been launched by a human with a prompt generated by a human or at a human's behest?
We haven't suddenly created machine free will here. Nor has any of the software we've fielded done anything that didn't originally come from some instruction we've added.
No. There are a countless other ways, not involving AI, that you could effect an email being sent to Rob Pike. No one is responsible, without qualifiers, but the people who are running the AI software. No asterisks on accountability.
Vinyl is having a resurgence. CDs are not.
Vinyl sale revenue is now 4.5x higher than CD sales, and CD sales are still dropping: https://www.riaa.com/reports/2025-mid-year-music-industry-re...
reply