Daily Digest

Sunday, 23 August 2026

This dated roundup collects the most interesting AI and technology developments found for Sunday, 23 August 2026.

41 min read • Published from the latest available digest

Policy & Ethics

The Guardian

An AI ‘debt bomb’ crisis? No. This isn’t Enron 2.0 | Gene Marks

Fears of a datacenter buildout debt crisis are exaggerated. The risks are different than in the past and they are recoverable Some experts are warning of a looming “debt bomb” crisis because big datacenter builders such as Meta , Oracle , xAI and CoreWeave are not only raising billions to construct these facilities but are also not recognizing these long-term debt obligations on their balance sheets. Should we be concerned? No. I know because I’ve seen this movie before. Continue reading...

Read more
marktechpost

The Developer’s Guide to NeMo Guardrails for Enterprise AI Safety

In this tutorial, we explore how to design production-grade safety for LLM-based applications using the NeMo Guardrails framework. We move beyond simple prompt filtering to implement a layered architecture, featuring deterministic PII redaction, retrieval filtering, output masking, and policy-based tool gating. By integrating stateful multi-turn evaluation and detailed activation tracing, we demonstrate how to build an auditable, secure, and cost-effective AI assistant capable of managing sensitive financial interactions while maintaining strict compliance standards The post The Developer’s Guide to NeMo Guardrails for Enterprise AI Safety appeared first on MarkTechPost .

Read more
lesswrong

5 Things I Learned About People From Doing Stand-Up Comedy

(People had complained about me just posting a preview of the linked post, so here's the whole thing!) I’ve been doing stand-up comedy for two months now, which is a total gear shift from my previous job in AI policy. I’ve learned some fascinating things about how humor works, but also about how people perceive each other, and how crowds can be surprisingly hive-minded. For those interested in improving their models of people and the world, here are five of those things! 1. People need to put you in a box The first thing a stand-up audience craves, consciously or unconsciously, is for you to tell them what kind of person you are. Are you an artsy lesbian, a hard-working immigrant, a married average Joe? For some reason, people can’t laugh until they have a box to put you in. If they can’t place a comedian, half the audience’s minds will be occupied trying to figure out who is talking to them. Must be some sort of innate human drive to quickly categorize new people you meet. This is why so many stand-ups start by addressing their appearance, accent, etc. (“I know what you’re thinking… [generic punchline about their appearance].”) I experienced this need to categorize myself when watching  this set by Robby Hoffman  recently. I couldn’t quite figure out who was speaking to me and just felt like I couldn’t quite settle until I got more of a grasp on who she is as a person. (This is less applicable to comedians who don’t tell stories or talk conversationally, and instead purely do one-liners. This style of comedy can feel less personal, so I suppose that’s why the urge to know more about the comedian may be lower.) In comedy, something glaring in the audience’s mind that goes unaddressed is called a “loop”. Common loops are a strange/loud laugh in the crowd, something going wrong on stage, or mentioning something that makes the audience concerned, such as the death of a relative. Generally, comedians want to “close” such loops as quickly as possible because people just can’t laugh while thinking “She just knocked over the microphone and hasn’t addressed it!” or “Is he okay?” Some comedians go as far as to address their appearance/voice/etc. every time. Helen Bauer says she has to address being fat every time she goes on stage because there are some people in the audience to whom apparently this poses a loop. Some people’s loops are dumb. I think people’s need to categorize comedians as soon as they get on stage is basically a kind of loop. And I think the fact that people do this to comedians is indicative of people doing this in everyday life too. 2. Laughter is very un-individual I had NO idea how un-independent people’s impulse to laugh is of the people surrounding them until I got into comedy. You will get flat responses, awkward silences, crickets one day and then roaring laughter the next  with the exact same set . And it’s not a case of a few people laughing the first day while most stay silent, and the reverse the next day. People’s laughter is pretty even on a given day; people  adjust  to the energy in the room. (What determines the level of energy in a room, other than the quality of the stand-up set? Probably some combination of the energy level right before the set and the presence of a couple of nutters who unconsciously  do not  gauge whether and how loud other people are laughing before laughing themselves. Yay, thank you, nutters.) The thing is, most of us do not listen to a joke, evaluate it on its own terms, and decide to laugh. Instead, we are extremely sensitive to how much other people around us are laughing. If the energy is flat, we too are flat. The crazy thing is we don’t  choose  to laugh more or less. We simply find it more or less funny. The same person could have a completely different experience at a show, depending on the night they choose to go. And they would never know! 3. Different countries’ audiences differ I suppose many people would expect different countries to have different affinities for stand-up—polite laughter in Britain, raucous crowds in Latin America, frigid Germans. I’m generally quite skeptical of such country stereotyping. But for stand-up audiences, it’s true. I can speak best to the difference between British and German audiences since I’m German—it’s huge. You could watch the funniest German stand-up clip you have ever seen on Youtube, and the audience will be giving nothing. ( Little treat if you speak German. ) Stand-up isn’t big in Germany at all; people aren’t that into it. British audiences are super nice in comparison! The strange thing is, I  totally get  why Germans are like that.  I would be like that  if I hadn’t moved to the UK years ago. It’s a cultural difference in how you interact with strangers. Germans don’t talk to random people in queues or at bars, etc., in the same way that British people do. Germans don’t cheer easily or join in some sort of audience/crowdwork interaction. To a German, being that outgoing can seem fake and tacky, not to mention exhausting. (It can feel like Brits have a bit more goodwill towards their fellow woman, and honestly, having lived here for a while, maybe Germans could use more of that!) I hear that Belgian audiences are similarly tough, and I imagine the Nordics might be even worse (or maybe that’s just my stereotypes). Reputedly, black audiences also differ from mainly-white ones, but I can’t speak to that difference myself. 4. There’s no single way to work From my background in AI policy, I’m used to there being quite a clear model image for how to become successful in a given line of work. You want to be a researcher? Get yourself a good supervisor, produce a lot of papers, and network. Want to work in policy? Move to London (or DC), go to a million happy hours, and learn to squeeze yourself into really short attention spans. Most importantly, work a 9-7 workday Monday to Friday, or ideally, Monday to Sunday. I’m sort of shocked at how there seems to be absolutely no theme to how successful stand-up comedians work. Not only is no one doing 9-5s, they don’t even agree on what writing looks like at a very basic level. Some comedians sit down with a pen and paper in an office, others compose lines in their head as they go about their day (wtf, how?), and finally, some nutcases  go on stage  with nothing more than an idea and just figure it out (??). Much in contrast to the ethos in the professional worlds I know, there just doesn’t seem to be an agreed-upon model way to become successful. To give some examples,  Nate Bargatze , one of the most successful US comedians, apparently never sits down with pen and paper. Instead, he composes his sets entirely in his head, without even writing them down. (He said this on  an episode of the Good One podcast , if I remember correctly.) Judy Carter, author of the “Comedy Bible”, recommends ranting out loud about a topic to find jokes, which doesn’t click for me at all but works for many people. Many experienced comedians go on stage with little more than an idea and just riff (e.g.  James Acaster  does this). This is sometimes called “writing on stage”. Some comedians like Chris Rock have word-for-word scripts and choreograph their act down to the raise of an eyebrow.  Maria Bamford  practices her sets standing directly in front of a mirror. Others have the key notes in their head and fill in the rest. My general sense regarding work ethic is that UK and Australian comedians are a bit laissez-faire about it. At the same time, comedians in New York (the stand-up capital of the US) will often do 2-3 shows  a night . The most “sensible” comedian I know of is probably Jimmy Carr, who treats it like a 9-5 job (though surely complicated a lot by touring etc.) I suppose what we call stand-up comedy is really a blanket term for quite a range of formats and styles. It encompasses storytelling, crowdwork and some improv, tightly scripted one-liners, characters, clowning, etc. It makes sense that the “process” differs for these. Stand-up writing is also such a personal thing, unlike, e.g., writing a research paper, so again it makes sense that the process is quite personal. The only agreed-upon parts of the success formula are, as far as I can tell, get as much stage time as possible, and be nice. Received wisdom is that doing a lot of gigs makes you better (though it’s a bit unclear to me how that trades off against writing time) and that assholes don’t get booked. You also hear many stories about people regretting not being nicer to the random MC who went on to produce [insert famous TV show here]. 5. People have different senses of humor Duh. But still, I did not fully appreciate this until I got into stand-up. People will be doubling over at sets of one-liners that will leave me absolutely cold. I don’t have much of an appreciation for clever construction of “joke” jokes, apparently. On the other hand, I will be in tears watching  Sam Campbell get a hall full of people to bow to images of apes , while other people are simply confused. I’m often completely unable to tell whether a stand-up set is good or not, because it’s just not my sense of humor; it’s not  for me . (E.g. I know  Maria Bamford  is the GOAT, but I will never understand why.) Amy Matthews articulated it something like this  on the Always Be Comedy podcast : We don’t have the same cultural fluency about different genres and tastes of comedy that we do with music or movies. If you’re not into heavy metal, you wouldn’t go to a heavy metal concert and then conclude “these musicians are shit”. But it is quite common for people to conclude “this comedian is shit” when the comedian is simply not their sense of humor. We don’t even have good names for different senses of humor, which goes to show how underappreciated these divisions in taste are. This is all just to say, if you ever see  me  do a set, please resist the urge (you’ll no doubt feel) to shout “You’re shit!”, and instead shout “You’re amazing at a style of comedy that just isn’t my cup of tea!” Thank you. If you liked this post, please check out my substack ! Discuss

Read more

Industry

forum

Moral Imagination & Effective Altruism

Published on August 23, 2026 3:47 PM GMT Michael Nielsen has a beautiful new essay on moral imagination : the ability humans have to 'develop and transmit new notions of good action, indeed, even new kinds of good'. As examples, he gives: Hammurabi's creation of a code of laws (in 1754 BCE) and his justification of his rule not in terms of divine or hereditary right, but by the delivery of justice to his citizens. St Gregory of Nyssa's argument (in 379 CE) that the entire institution of slavery was morally wrong. Joyce Roush's (1998 CE) idea that she should altruistically donate a kidney to help save the life of an unknown stranger.  Nielsen acknowledges that any particular act of moral imagination often builds on earlier precursors, but that this shouldn't lessen our appreciation for these major steps forward. One could think of moral imagination as a key part of making moral progress. Moral progress (moving towards better systems of morality) involves the creation of new approaches to morality, and the critiquing of what has been suggested. Through the iteration of creation and critique, we slowly make progress. Some of this process happens within moral philosophy (especially the critique), but much of it happens in the wider world. I see moral imagination as a key part of the creation step — allowing us to make surprising new changes to how we see morality. It probably isn't the whole of the creation step, as one can also use more systematic approaches, such as looking for cases where our current practice is morally inconsistent. In his essay, Nielsen mentions effective altruism by way of contrast: The moral imagination viewpoint contrasts with a picture in which our goal is choosing the best possible action from an existing menu. Choose the right charity to donate to, the right career, the right cause. This is a common everyday view, with Effective Altruism an especially well-developed instance, often concentrating on choosing between existing ways of doing good – is it better to donate a dollar to cause A, B, C, or D, on the GiveWell list of recommended charities? That is a valuable project. But the intellectual focus is on selection from the existing menu, which fundamentally underrates imagination 5 . When we can expand what it means to be good, contribution may lie not only in selection but in imagination — a fundamental virtue, expanding the range of good actions available to be selected from. I find myself tantalized (and sometimes worried) by this: what notions of good have we not yet imagined? What will we one day judge as evil that today we do not? I think this is correct as far as it goes, and I can see why choosing among a pre-existing list of moral options wouldn’t count as moral imagination. However,  effective altruism has created a surprising wealth of expansions of our moral concepts, self-conceptions, orientations to the world, techniques, and values.  Nielsen doesn't necessarily disagree, for in a footnote he clarifies: I am not saying Effective Altruism doesn't encourage and enable moral imagination. Rather it's not a primary focus (and is, I believe, underrated). Moral imagination probably is underrated in effective altruism (as it is almost everywhere). Indeed, I think even within EA, people are unaware of many of the ways effective altruism has contributed new things to moral thought. So I want to draw out some of the most significant examples. The idea of making a moral choice about where to donate based on the cost-effectiveness of the options is fairly obvious in retrospect. After all, if you are donating a particular amount of money and cost-effectiveness is the amount of benefit per unit of money, then choosing based on cost-effectiveness is really just choosing based on how much impact will be achieved. But this wasn’t being done and was resisted strongly by much of the philanthropy and ethics communities. Cost-effectiveness was seen as a tool for economists or for optimising one's personal spending. It wasn’t seen as something that morally serious people need concern themselves with. I think we challenged that quite successfully and thereby expanded the set of moral tools and considerations.  This focus on cost-effectiveness also led to ideas such as that where you give can be even more important than whether you give (e.g. switching from the charity that will save 10 lives with your donations to the one that saves 100 is a bigger difference than whether you donate to the one that saves 10 lives in the first place). And many previous groups that had focused on the morality of personal donation were often too focused on the donor (e.g. a vow of poverty to rid yourself of the evils of money, or the focus on making sure to be anonymous in your giving). In contrast, we centred the recipients rather than the donor. This helped one see that whether someone ultimately saves 10 lives or 100 is a much bigger moral issue than whether they gave anonymously. After all, the potential recipients would care vastly more about the fact that 10 times as many of them would be saved than about whether the donation was anonymous. The idea that we should apply the standards of scientific evidence to choosing a charity is also quite new. It is obvious in retrospect that when so many lives are at stake, we should apply humanity's best methods for determining what works, but people resisted this. In part I think this is because they felt that science and ethics are different magisteria (there was a similar reluctance, even revulsion, about using numbers in making moral choices). Of course, using scientific evidence to determine what to fund isn’t new from the perspective of the government departments doing international development work, but it wasn’t being used as a moral principle or part of individual choices. People found the idea that a morally good person should look at scientific evidence before donating (or seek the advice of others who have done so) to be surprising.  Another new concept is cause-neutrality — the idea of just trying to do good (impartially considered) and being open to completely changing where you give or what you work on based on which approaches are shown to be more cost-effective. Most people form strong moral motivations through concrete specifics — they are viscerally moved by witnessing some injustice or tragedy and are then motivated to help fight that kind of injustice or prevent that kind of tragedy. Or they form a strong attachment to a person or group and then make great sacrifices to defend the rights and interests of that person or that group. Cause-neutrality is saying that when trying to do good, we should be more impartial, so that we are attentive to situations when we have the opportunity to achieve much more in a different field or for a different group than those we’ve formed a passion around. Another new idea was that we should take seriously the fact that each of us can save hundreds people's lives with our donations, and that positive moral action was therefore a key part of our own lives (this built on Singer’s argument in 'Famine Affluence and Morality'). What we can do positively for others is often vastly more important than the intrinsic value of own lives and personal projects (because we can save hundreds of other people’s lives and personal projects…) and it is more important than many of the wrongs we commit in our day to day lives.  I therefore argued that our charity has central moral importance in living a good life. Indeed I originally called effective altruism ‘positive ethics’ because I was amazed at how much good we can actively do in the world, while moral philosophers (and religious ethics) is almost entirely focused on avoiding wronging others or doing bad things. We can do startling amounts of good, but no-one was talking about it. They praise a saint, but people weren’t asking regular people why they weren’t saints when it would be so easy to do so (at least among those with middle class careers in rich countries).  Moral uncertainty is also new. There was almost nothing written in moral philosophy about making decisions when you are uncertain about which moral theory is right (or are uncertain about the moral considerations that relate to the issue at hand). There was a lot on what to do when uncertain about descriptive facts, but not for when uncertain about morality, despite this being the puzzling situation we all find ourselves in. Will and I did a lot to explore the key similarities and differences with descriptive uncertainty. In the process, we became a lot more comfortable with our own moral uncertainty — acknowledging it and acting under it. This led to ideas like the value of moral information (which helps explain the value of doing moral philosophy), being aware of our moral ignorance and recommending we adopt a stance of moral humility . And of course one of the most distinctive things about effective altruism is that one of the principal cause areas we work on is protecting humanity from existential risks. This wasn’t something that was conceptually available before the year 2000. While effective altruism didn’t invent the idea of extinction risk or existential risk, we saw its vast importance and did a great deal towards explicating the moral considerations of this grand new area of moral action — understanding why it would be such a great tragedy in ways that connect with many moral traditions, and helping to take saving the world from sounding like a science-fiction story (or comic book plot) to seeing safeguarding humanity’s future as a fundamentally wise course of action for any species that has become capable of destroying itself. And connected to this is the idea of longtermism — that future generations matter deeply, and so since there could easily be tens of thousands more generations to come, ways of having lasting impact over all those generations are especially morally significant. Avoiding existential catastrophes is one obvious example, but avoiding other irrevocable losses counts too, as does doing good things that are very unlikely to be undone (such as ending slavery). Some of this had been seen by environmentalists (a similar argument explains why ecosystem destruction and species extinction are so bad) by economists (thinking about the power of economic growth over thousands of years) and others too. But the idea of seeing future generations as not just a little bit important, but deeply important — probably more important than all our own lives and projects — is a moral innovation. So I think effective altruism has done quite a lot of moral innovation or moral creation. And I think some of these examples would count as moral imagination too. That said, I think Nielsen is probably right that moral imagination being undervalued by many EAs — despite it pervading many of the concepts and ideas of the movement. Discuss

Read more
lesswrong

Prompt Sufficiency: A Missive for the Managerial Class

AI Assisted Work: A Missive For the Managerial Class The Smart Employee “ It doesn't make sense to hire smart people and then tell them what to do, We hire smart people so they can tell us what to do.” - Steve Jobs This famous quote is bound to provoke some discomfort: What is the meaning of leadership if it dictates neither what nor how ? The answer, according to Mr. Jobs, is that effective leadership is sometimes accomplished through yielding control, rather than holding on to it. By delegating aspects of the what , we are able to leverage the experience of others. This principle is an acknowledgment of the human limitations of the leader - no matter how capable - and a strong endorsement of the value of teamwork, humility, expertise, and thrift! After all, hiring capable people (read that as: expensive people) and reducing them to automatons is a tremendous waste. On the topic of waste, it has been widely noted that large gains in productivity from Artificial Intelligence have been slower to materialize than expected. Having utilized LLMs quite effectively, and having carefully observed others stumble, I have a great deal of conviction that the gross misapplication of the above leadership principle is largely to blame. The problem is actually quite simple: we treat LLMs like smart employees, but that is not precisely what they are, appearances notwithstanding. It is becoming apparent that the wisdom of Mr. Jobs simply does not apply here. To make the point a bit clearer, I think it is helpful here to meditate on what a smart employee actually does, because “smart” in the purely academic sense doesn’t quite fit. The opening to the famous 1899 essay “Carry a message to Garcia” by Elbert Hubbard is a bit closer - it is one who can be entrusted with accomplishing a task with little fuss. Someone said to the President, “There is a fellow by the name of Rowan [Lt. Andrew S. Rowan, U.S. Army] will find Garcia for you, if anybody can.” Rowan was sent for and given a letter to be delivered to Garcia. How the “fellow by the name of Rowan” took the letter, sealed it up in an oilskin pouch, strapped it over his heart, in four days landed by night off the coast of Cuba from an open boat, disap­peared into the jungle, and in three weeks came out on the other side of the Island, having trav­ersed a hostile country on foot, and delivered his letter to Garcia—are things I have no special de­sire now to tell in detail. The point that I wish to make is this: McKinley gave Rowan a letter to be delivered to Garcia; Rowan took the letter and did not ask, “Where is he at?” By the Eternal! there is a man whose form should be cast in deathless bronze and the statue placed in every college of the land. It is not book-learning young men need, nor instruction about this and that, but a stiffening of the vertebrae which will cause them to be loyal to a trust, to act promptly, concentrate their ener­gies: do the thing—”Carry a mes­sage to Garcia.” Raw cognitive ability here is pure table stakes. It is the intellect which makes competence possible; it is the ability to “fill in the details” which makes the smart employee a true operator - a worthy trustee of the what . Rowan knew of the necessity of a map and therefore acquired one, understood himself well enough to know if any refresher of trail-craft was needed, anticipated obstacles and prepared backup plans, and so on. One could delegate the what to such an employee, but such an employee could easily execute tasks as well. The valuable activity here is gap-filling: working out the details that lie in the jungle between the letter and the general. The human being knows when it has enough information; the human being knows when it does not. It knows when it is safe to guess. It knows when guessing is dangerous. All these abilities are brought to bear to create the “experience” of the effective report - a clean, no-fuss operator. Interacting with Agentic AI creates the impression that it is that kind of employee because it carries so many of the same signifiers: breadth, encyclopedic depth, speed, and some indicators of taste and judgment. This indicates intelligence - which, after all, is a common trait amongst productive people. But here’s the reality: machine intelligence - in its current incarnation - must be utilized , not unleashed. My precise claim is that there are two points of failure that are compounding, both of which are managerial rather than technological. The first is that our prior experiences with direct reports have made us blind to the sheer amount of information - or context - required for successful execution of anything non-trivial. The second is that LLMs are essentially starved of this context by the narrow band of communication available to them: the prompt. The remedy to both problems is prompt sufficiency . I’ll develop this concept in more detail shortly. Here’s the crux: The context we are accustomed to human beings acquiring independently and without explicit instruction to do so is not automatically acquired by AI Agents. We have become so accustomed to this invisible human activity that we have become blind to its necessity, and surprised when failure to provide necessary context results in failure. An Example Here is an example: Please digest the last five quarterly earnings reports and write a memo suggesting a strategy which addresses any shortfalls you see in our current approach. This is a perfectly reasonable request for a capable employee, but there are several unstated yet load-bearing assumptions. For example: What is the organization’s capital risk-management philosophy? This is unlikely to be spelled out in the quarterlies, but is likely to influence what is considered an actual shortfall in strategy versus an acceptable risk. The Agent AI’s report will appear to be well-thought-out and thorough. It will reference the appropriate figures. We take this as a signifier of the qualities that a strategy memo ought to have: diligence, thoughtfulness, thoroughness. But those are only signifiers: is it truly fit-for-purpose? Here is another example: “ Please digest the last five quarterly earnings reports and write a memo suggesting a strategy which addresses any shortfalls you see in our current approach. Bear in mind that we consider any indications of a future liquidity crunch to be unacceptable, and utilizing available credit lines is an acceptable recourse as long as recurring revenue remains above … ” Here we are providing a great deal more of the context required to produce a result which is fit for the situation at hand. There was no way that the AI could have acquired this additional information from sources like general knowledge because it pertains to how the executive responds to risk. Prompt Sufficiency A prompt is sufficient when an agent is able to complete the work product without arbitrary choice. Sufficiency is not a qualitative property - it is not a function of format, appearance, taste or style. The presence of imperatives (”Make no mistakes!”), role-playing directives (”You are a terraform engineer”), or descriptors (”thoroughly…”) do not add to or take away from sufficiency. Sufficiency is related to the idea of information content - it is a combination of deducibility and the absence of a requirement for mind-reading on the part of the machine. Given the prompt and the available information, is the agent required to know something that only you would know - assuming, of course, that you have an understanding of what needs to be done? Or, must the agent be unduly creative, clairvoyant, or decisive on your behalf? If so, the prompt is probably insufficient. Here are examples of insufficient prompts: Design an appealing dashboard (Is it possible to deduce whether a dashboard will be appealing to you without reading your mind?) Create a new strategy for revenue growth based on the last five quarters' financials (Have you expressed a criterion for what you consider to be an acceptable strategy? Or is this known only to yourself?) It is possible to develop a nose for insufficiency by making a practice of mentally elaborating on the task at hand. Suppose you are a stranger (assume you have expert background in whatever topic), and that you are asked to complete the work. Here is the question: Is it possible to deduce how to complete the work with only expert domain knowledge and the prompt? Or are there particulars to the task that cannot be deduced from subject matter expertise? As the asker, ask yourself: What needs to be specified that is particular to the task? Let’s compare two prompts: “Build a login screen using the brand kit I provided earlier”. “Build a login screen using the brand kit I provided earlier, with a visual design matching the most popular app on the App Store”. Despite neither prompt containing much detail, the second prompt is sufficient; the first is not. Why? This is an exercise for the reader. Discuss

Read more

Research & Products

simonwillison

llm 0.33

Release: llm 0.33 My highlights from this release: Upgraded to the OpenAI Python library 3.x and switched the HTTP client dependency from httpx to httpx2 . #1608 , #1631 I shipped a quick 0.32.1 fix for this yesterday, but this is the more comprehensive fix. llm embed and llm embed-multi now accept --key . The Python EmbeddingModel.embed() , EmbeddingModel.embedmulti() , Collection.embed() and Collection.embedmulti() methods accept key= too, passing the resolved per-call key to embedding plugins without changing shared model state. Existing plugins that read self.key continue to work through a compatibility fallback. Thanks, ChrisJr404 . #757 , #1620 The embedding models now use the same pattern for keys that regular LLM models do. llm prompt -t/--template can now be repeated to combine templates in order. This allows model configuration and options from one template to be used with a prompt from another. This unlocks a neat pattern where you can create templates that package a model with a set of default options: llm -m gpt-5.6-luna -o reasoningeffort high --save lhigh llm "Generate an SVG of a pelican riding a bicycle" --save pelican # Combine and run the templates llm -t lhigh -t pelican Reasoning-capable Responses API models now support a reasoningsummary option with auto , concise , and detailed values. This can be used with llm openai endpoint --responses . #1600 This is particularly useful for exercising different models that provide their own imitation of the OpenAI Responses API. Tags: annotated-release-notes , llm

Read more
thenextweb

Harvey’s first in-house model for legal work is here

Tenet is post-trained on Moonshot's Kimi K3, which means an OpenAI-backed legal software company has put Chinese open weights underneath privileged client work Harvey has launched Tenet, its first proprietary model for legal work, post-trained on Kimi K3, the open-weight model from Chinese startup …

Read more

Tags: ai_safety, announcement, opinion

lesswrong

When would you leave Anthropic? Notes from a chat with a capabilities researcher

I was at a house party hosted by an AI Safety friend of mine. I join a conversation midway, where my friend is saying that doing capabilities research at frontier lab is evil, given the catastrophic risks. Nothing out of the ordinary, until I find out the person who they are talking to is a capabilities researcher at Anthropic! I was shocked at the directness of my friend. The researcher took it well though, partly based on their temperament, and partly because they have been exposed to several AI Safety spaces previously. Unfortunately, at this point, my memory of the conversations is extremely shaky. Also, the conversation was not continuous and happened piecemeal between other ongoing discussions. But here is my best recollection. The researcher responds with something like: “Not a good use of our time to go flesh out our positions, as we will just re-hash the standard arguments. For example, one counter is that Anthropic is creating products that people pay for and find valuable. Or, [insert another tepid argument that I can’t remember]”. I was surprised that this is the first thing they thought of, rather than talking about the potential gigantic upsides of advanced AI systems. As much as I lean libertarian, justifying catastrophic risks with ‘customers pay for our products’ was pretty weak, at least with the particular way they phrased it. I have some 1-1 discussion with the researcher, and sympathise with them being called evil. I bring up vague idea I have had that researchers in frontier labs should have some kind of public statement about personal redlines: what things would AI’s or Anthropic or Anthropic leadership do that would cause them to leave the company. They respond saying that this would likely just be a checkbox exercise, with people copying and pasting some standard meaningless statement which has no teeth or consequences. I say that instead, maybe people should just post a statement every six months along the lines of “I have reflected on the risks and benefits of working at [frontier lab], and have decided to [stay/leave]”. Sure, again, they could just copy and paste this and use it as a checkbox exercise, but I sense that most people would not post such a statement if they had not actually done a reflection, whereas people might honestly post a statement about their personal red lines, and then change their minds about it later. After some discussion, the idea morphed to: every six months, they have a discussion with somebody – e.g. me – where the aim is to help them explore their own personal views. They disliked framing, saying it would be adversarial and feel like an inquisition with lots of AI safety people around challenging them. I said the discussion would be 1-1, would be done on whatever terms the other person wanted, they could leave the discussion when they wanted, and that my personal style is to just try to understand the other person, rather than to explicitly change their mind. They then started asking a question as a counter, “What would you think about having a discussion every six months about big things in your…”. They did not finish the question, because the answer was evident to both of us. Yes, that would be useful! Of course it would. At this stage that they had no good reason not to do this exercise, which of course is separate from them actually doing it. The final question I ask – in an attempt to identify an underlying crux between us – was, “Do you think it is possible, in the next 10 or 20 or 40 years, for an AI system to be created that causes human extinction?” They responded immediately, “Yes, of course.” I can only assume that the expression on my face was asking: “So why are you working at Anthropic then?!” It was interesting to see the cogs turning in their head head realtime: I am confident that there is a significant part of them that believes they should not work at Anthropic, and that the other parts of them are desperately trying and failing to come up a coherent reason not to think about the issue. As we part ways, I say that the offer to have this discussion is real, and they are welcome to stay in touch. I sense this is unlikely to happen, with the fact we did not share contact details being the least relevant reason. Final thoughts. I am curious to know what you think. I believe that direct outreach to capabilities researchers is a useful lever, at least if done respectfully and collaboratively. Going up to them and calling them evil to their face might nudge some people, but I think much less so than having them think things through for themselves. However, I have no clue how this would actually work logistically or how to make something like this happen. EDIT: My friend sent me this clarification: “I didn’t literally call them evil. My sentence was closer to ‘I think people who work for frontier labs are committing an evil act, but that’s weird, because I don’t think these people are evil’. I agree that calling people evil doesn’t really change people’s minds.” I believe that frontier lab employees – especially those that are catastrophic-risk-pilled or those who signed the Pacing the Frontier open letter – are obliged to regularly and publicly share some kind of statement regarding their red lines and whether they would leave. And then ideally following through when those redlines are crossed. The reception to Alex Turner’s post about why they left GDM is indicative of the value of such actions. I believe that people more broadly should have personal redlines regarding the use of frontier AI systems. There is something incoherent about believing a a company might cause human extinction, but then also being their customer. This is something I still need to think about for myself. Discuss

Read more

Tags: hardware, reasoning

lesswrong

Twenty Years from RSI to Takeoff: Slow Learning, Scaling Slowdown, Industrial Explosion

Industrial explosion is what will make the next-model building loops (and thus learning) with LLMs 1000x faster by about 2050, if indeed the slow-learning prosaic RSI becomes AGI before the big compute buildout slowdown of 2032+ that is already starting. This puts an upper bound on how long it takes to invent ASI that sets off software-only singularity, implementing efficient online learning and fixing all the other hobblings of the likely near-future AGI technology (LLMs/pretraining/RL). The invention of ASI in that sense is still possible at any time (and very quickly scales, given all the compute), but the likely initial state of slow-learning AGIs of 2028-2032 doesn't seem to give them a significant advantage over humanity in getting there faster. And so it doesn't seem too unlikely that nothing substantively new gets invented until 2040-2050, when the LLM/RL AGIs start accelerating because of the industrial explosion they set off. Fast Reasoning, Slow Learning The current methods are likely to enable automated general learning (thus AGI) very soon, using automated creation of RL tasks/environments/graders filling the visible gaps in model capability for the topics and situations that happen to be borderline unfamiliar for that model, followed by automated next-model building. This teaches LLMs deep skills, but operates at the speed of next-model building (weeks to months for one iteration of advancing the deep skill frontier) rather than at the speed of next-token generation (100-1000x the human speed). Humans are not the key bottleneck to the speed of next-model building loops, there's still a lot of waiting for the compute to do its thing in training, so achieving prosaic RSI by teaching the near-future LLMs all the skills necessary to perform it automatically doesn't make it go too fast. Using smaller LLMs to make everything faster doesn't work because the current frontier LLMs are probably borderline insufficient for learning the prosaic RSI skills that automate the next-model building loop. The LLMs of 2028-2031 (that are very likely sufficient) will be even bigger, though the cost of training or running them is not as bad as "quadrillion total params" sounds . This cost can't be circumvented by using different hardware that makes LLM inference much faster, because different hardware doesn't reduce the necessary number of FLOPs, which are not terribly wasted even in RL training and inference that involve bandwidth bound decode. Fundamentally, cost is the amount of compute, and the only thing that overcomes it is the scale of the global buildout. Compute Slowdown, Industrial Explosion Since LLMs become ready to close the next-model building loop (that enables AGI) just as the human industry runs out of various kinds of fuel for quickly increasing the scale of the compute buildout, there is no opportunity for another near-term 1000x increase in the available compute (and thus the speed of next-model building loops), the way compute was increasing in 2022-2027, and the way it'll keep increasing (a bit slower) in 2028-2032 until the pace of decommissioning old compute somewhat catches up to the 2028+ pace of producing new compute, set by factors like availability of EUV machines and skilled human labor. Thus the big compute slowdown of 2032+, stronger than the end of the current exponential scaling of compute by 2028+. Without paradigm-breaking algorithmic innovations, the slow-learning AGIs can't quickly invent such innovations, and so the more predictable component of the pace of progress is set by the pace of the compute buildout. But also, the AGIs (likely available since 2028-2032) make the industrial explosion of robot-building robots a predictable medium-term development. It probably doesn't start right away, since the AGIs are not much faster than humanity at taking care of all the novel engineering challenges (requiring many next-model building loops to get good), and before it goes into full swing humanity still needs to handle the industrial side of things (at the human pace). The automotive industry and the compute buildout acceleration of 2022-2027 seem like good anchors for how this might unfold. The process starts once AGIs unlock an outsized demand for robots (by making them very useful for everything), and the industry starts reshaping itself to increase the supply as fast as it can, similarly to the consequences of the ChatGPT moment. Robot production exhausts the industrial capacity of the supply chains within 3-5 years (similarly to how it took 5-6 years to reshape compute production). At that point, the scale of the robot supply (anchored to the current automotive industry) approaches a significant portion of human labor, so the process continues right past the limits of human industry without another big slowdown. If the doubling time of the robot-building industry (autonomously operated by AGIs using the existing robots) is around 1 year, then 10 years of this process increase the industrial capacity about 1000x. The countdown should probably start from the end of the 3-5 year period of industrial conversion, when the robot industry first matches a sufficient fraction of the human industry to also start producing as much compute (together with all the other precursors), and the slow-learning AGIs probably need the time for the next-model building cycles to figure out how to automate everything. Prosaic Timeline to Takeoff The timeline starts with prosaic RSI in 2028-2032. The resulting slow-learning AGIs first make robots very useful generally within 2-3 years, in 2030-2034, setting off the industrial conversion of human industry towards robot production. This lasts another 3-5 years, and by 2034-2039 the automated robot-building industry operated by robots and AGIs first matches the human industry's capacity in terms of the compute buildout it can support. If this industry can quickly reach a doubling time of 1 year, it can 1000x the compute buildout by 2044-2050. At that point, the next-model building cycles take 1000x less time, and so the unpredictable paradigm-breaking algorithmic innovations necessary to set off a software-only singularity happen on the scale of months instead of centuries, and would've happened at some point earlier than that, probably by 2040-2045, but certainly by 2050. This is the upper bound on the timing of feasibility of superintelligence, which mostly assumes just the current paradigm (extremely hobbled in its efficacy at superhuman invention) rather than any particular future breakthroughs. Discuss

Read more

This digest was automatically generated • 41 min read

ainewsdigesthardwareproduct_launchannouncementfundingregulationreasoningopinionindustry_movemodel_releaseai_agentsacquisitionai_safetypolicyopen_source