AI News Archive: July 27, 2026 — Part 7
Sourced from 500+ daily AI sources, scored by relevance.
- AI-native Canadian startups outpacing international peers, AWS finds
Report credits deep technical expertise, research talent, and founders used to doing more with less. The post AI-native Canadian startups outpacing international peers, AWS finds first appeared on BetaKit .
Score: 28🌐 MovesJul 27, 2026https://betakit.com/ai-native-canadian-startups-outpacing-international-peers-aws-finds/ - How AI-Powered Risk Engines Are Redefining Fraud Prevention for NBFCs & Lenders
For years, the story of fraud in Indian banking was mosty a card and internet-transaction story: high volume, relatively low value, chipped away at steadily through OTPs, device binding, and better authentication. That story just changed. The Reserve Bank of India’s Annual Report for 2025-26 shows fraud cases involving Rs 48,021 crore, up from Rs […] The post How AI-Powered Risk Engines Are Redefining Fraud Prevention for NBFCs & Lenders appeared first on CXOToday.com .
- DevRev: Enterprise AI has a goldfish problem (and it is on your invoice)
DevRev: Enterprise AI has a goldfish problem (and it is on your invoice) Gartner
- I Tested Every GPT-5.6 Feature — These 10 Are the Real Game-Changer
Most people are only using a fraction of GPT-5.6. These 10 features unlock its real power. OpenAI’s newest model family launched under a genuinely strange cloud a government-vetted preview before the public even got a look. I spent three weeks past that rollout actually using it. Here’s what’s real and what’s noise. GPT-5.6 isn’t just another AI update, it’s a glimpse into the future of intelligent software. I’ll admit the timing of this release made me suspicious before I’d even opened the app. GPT-5.6 didn’t roll out the normal way. OpenAI first previewed it in late June to a small group of “trusted partners” whose participation had reportedly been shared with the U.S. government, after the administration asked the company to delay a full public rollout over cybersecurity concerns. Sam Altman later described the process as a “collaborative back and forth,” and OpenAI publicly said it didn’t want that kind of vetting process to become the norm going forward. Then, on July 9th, it opened up to everyone on a paid plan. That’s an unusual amount of political weather for a model release, and it made me want to actually verify the hype myself instead of trusting the launch-week headlines. I’ve been burned before by exactly this kind of release cycle a dramatic rollout story generates a wave of coverage, everyone forms an opinion off the headline, and by the time the actual model is boring and available to everyone, half the internet has already moved on without really testing it. So I spent three weeks running GPT-5.6 specifically the flagship Sol variant, with some side-by-side testing on Terra against real work: actual coding tasks, actual documents, actual design decisions I needed to make for a client project. Not benchmark screenshots. Real use. Ten features held up. Here they are, in the order I’d actually rank their impact on my day-to-day work, not the order OpenAI’s press release put them in. First, the Basics: What You’re Actually Choosing Between Before the list, it’s worth knowing what you’re picking, because GPT-5.6 isn’t one model it’s three, and picking the wrong one for your task wastes either money or quality. A side-by-side look at Sol, Terra, and Luna — the three pillars of the GPT-5.6 family. All three share a genuinely enormous 1.05-million-token context window with up to 128,000 tokens of output. I want to flag something that surprised me here: Terra, the mid-tier model, scored 87.4% on Terminal-Bench 2.1 against Sol’s 88.8% a gap so small that for most coding work, Terra is the smarter financial choice, not a compromise. I didn’t expect to end this test recommending the cheaper model for most tasks, but the numbers genuinely pushed me there. 1. Prompt Caching That Actually Saves Real Money This is the least glamorous feature on this list and the one that changed my monthly bill the most. GPT-5.6 introduced explicit cache breakpoints and a 30-minute minimum cache life, meaning repeated context a long system prompt, a big document you’re referencing across several messages gets billed at a fraction of the normal input rate. Cached input runs at roughly a tenth of the standard price. I was running a document-analysis workflow that re sent the same 40-page reference file across dozens of queries, and once I restructured my calls to actually use the caching properly, that portion of my bill dropped by more than half within the same week. The catch, and I learned this the annoying way: cache writes are billed slightly higher than normal input (about 1.25x), so if you’re not actually reusing that cached context multiple times, you can end up paying more, not less. I found this out after restructuring one workflow too aggressively, caching a document I only ended up referencing twice, and watching my per-query cost tick upward instead of down. This isn’t a “turn it on and forget it” feature. It rewards understanding your own usage pattern, and punishes you quietly if you don’t. 2. ChatGPT Work: An Agent That Actually Touches Your Files ChatGPT Work launched alongside GPT-5.6, and it’s the feature I was most skeptical of going in, mostly because “AI agent for work” has been an overused, underdelivering phrase for a couple of years now. It’s built to gather context across your connected apps and files and then actually produce something a document, a spreadsheet, a presentation rather than just describing what one should look like. I connected it to a folder of client research and asked it to build a first-draft competitive analysis. It pulled real numbers from the actual files I’d given it instead of inventing plausible-sounding placeholders, which is exactly the failure mode I was bracing for. It’s available across web, phone, and desktop, starting with the Mac and Windows apps. I still reviewed every number it pulled before sending anything to a client, because that’s just how I work now regardless of which tool produced the draft. But the first-draft quality was genuinely higher than what I’ve gotten from similar “agentic” tools over the past year. 3. Codex Folded Directly Into ChatGPT’s Desktop App If you’ve used OpenAI’s Codex tooling separately from ChatGPT, that separation is gone now. Codex is built directly into the ChatGPT desktop app on macOS and Windows, with a preference toggle to favor Codex-style behavior, including changing the app’s icon to reflect it. Practically, this meant I stopped context-switching between two separate apps for coding work and general assistance. I didn’t expect a UI consolidation to matter this much to my actual workflow, but removing that context switch measurably cut down how often I lost my train of thought mid-task. 4. Genuine “Design Judgment” in Computer-Use Tasks OpenAI specifically calls out improved design sensibility in GPT-5.6, the ability to take high-level direction and produce interfaces that are, in their words, tasteful and ergonomic, not just functional. The part that actually impressed me wasn’t the generation step. It was that the model’s stronger computer-use capability lets it inspect the rendered result afterward, not just the underlying code, and catch visual problems before handing the work back. I tested this by asking it to build a simple landing page with only a rough verbal description of the vibe I wanted. It generated the page, then flagged on its own that a button’s contrast ratio was too low against the background and fixed it without me asking. That’s a genuinely different category of behavior than generating code blind and hoping it renders correctly. 5. A Coding Model That Backs Up Its Claims OpenAI calls Sol its best coding model yet, and for once the benchmark table mostly supports the marketing. Sol leads the Coding Agent Index and Terminal-Bench 2.1 at 88.8%, and scores 64.6% on SWE-Bench Pro. On real repository work actual bug fixes in a mid-sized Python project I maintain it handled multi-file changes with fewer follow-up corrections than I needed with the previous generation of models I’d used for the same kind of task. It’s not universally ahead of every competitor on every chart. Anthropic’s models lead on some coding benchmarks by a wide margin depending on which one you check. The honest takeaway: Sol is genuinely excellent at agentic, terminal-heavy coding work specifically, which is a narrower and more useful claim than “best at coding” in general. There was one specific bug fix that stuck with me a race condition in a job queue that I’d been putting off dealing with for weeks because I knew tracing it manually would eat half a day. Sol found the actual root cause on its second attempt after I gave it access to the relevant logs, and explained the timing issue clearly enough that I understood why the fix worked, not just that it did. That’s the difference between a tool that patches symptoms and one that actually helps you understand your own system better. A snapshot of the GPT-5.6 features that matter most for developers, creators, and AI power users. 6. Real Token Efficiency Gains, Not Just Marketing Language Altman has publicly claimed Sol is 54% more token efficient on coding tasks compared to the previous generation. I can’t independently verify that exact figure, but I did track my own token usage across comparable coding sessions before and after switching, and my effective cost per completed task dropped noticeably not by exactly that percentage, but enough that it wasn’t just a rounding difference either. This matters more than a single flashy benchmark because it compounds. A model that’s slightly smarter but meaningfully cheaper per task changes how liberally you’re willing to use it day to day, which changes your actual workflow, not just your invoice. 7. A Cybersecurity Focus That’s Unusually Specific OpenAI is marketing GPT-5.6 as its strongest cybersecurity model yet, explicitly built for defensive work threat modeling, code review, patching, and blue-team-style analysis achieved with meaningfully fewer tokens than before. This is also, not coincidentally, the exact capability that triggered the government-vetted preview rollout in the first place. I don’t do offensive security work, so I tested this narrowly: I fed it a codebase with a handful of deliberately introduced vulnerabilities and asked for a security review. It caught the injected issues cleanly and explained the actual attack surface in plain language, which is the part that’s harder to get right than just flagging that “something looks unsafe here.” 8. Hosted Sites, Built Right Into the Product Alongside GPT-5.6, OpenAI added a hosted sites feature for paid users the ability to go from a generated page to something actually live and reachable by a URL, without leaving the product to configure hosting somewhere else. I used this for a quick single page project brief I needed a client to see immediately, and skipping the usual deploy-and-share dance saved me a genuinely annoying ten minutes I usually don’t think about until it’s missing. It’s not a replacement for real hosting infrastructure on anything that needs to scale. For quick, disposable, shareable pages, it’s a small convenience that adds up over a lot of small tasks. 9. Effort-Level Control That Actually Changes Behavior Paid users can select among Sol, Terra, and Luna and set a distinct effort level for each, rather than the model silently deciding how hard to think on your behalf. I used this deliberately across the three weeks low effort for quick lookups, higher effort for anything I was going to ship and the difference in response depth was noticeable enough to matter, not just a placebo setting buried in a menu. The part I appreciated most is that this puts a real cost-versus-depth decision back in the user’s hands instead of hiding it behind an “auto” toggle that quietly picks for you and shows up as a surprise on your bill later. 10. A Genuinely Massive, Genuinely Usable Context Window A 1.05-million-token context window sounds like a spec sheet number until you actually load something close to its limit and watch what happens. I fed it a full technical specification document alongside several related design docs well past what I’d normally consider “long context” and asked questions that required connecting details from the beginning and end of that combined material. It held together. Not flawlessly I noticed a couple of moments where a very early detail got slightly under-weighted compared to more recent content but meaningfully better than the context degradation I’m used to seeing once a conversation gets genuinely long. For anything involving large codebases or lengthy reference material, this is the feature that actually changes what kinds of tasks are practical to hand off at all. What I’d Actually Tell a Friend If you’re deciding whether to bother switching, here’s the honest short version: Terra is the variant most people should start with, not Sol. The price difference is real, the benchmark gap is small, and unless you’re doing genuinely demanding agentic coding work at scale, you likely won’t notice what you’re missing. The features that actually changed my daily workflow weren’t the flashiest ones from the launch keynote. It was the boring combination of cheaper caching, a coding assistant that stopped living in a separate app, and a context window that held together under real pressure instead of just looking good in a spec sheet. The design-judgment and cybersecurity capabilities are genuinely impressive, but for most people reading this, they’ll matter far less than the three unglamorous wins that quietly make everything else you do a little faster and a little cheaper. I went into this test annoyed at the rollout politics and expecting to write something more skeptical than this ended up being. I’m still not thrilled about a government vetted preview tier becoming a thing companies feel comfortable doing. But the model underneath all of that noise is genuinely good, and I think it’s worth separating those two facts instead of letting one color the other. If you’re a developer, creator, or AI power user, GPT-5.6 is worth exploring not because of the benchmark numbers, but because of how it changes your daily workflow. In the end, the best AI is the one that helps you ship better work, faster. Which of these ten would actually change your workflow, and which ones feel like launch-keynote filler to you? I’d like to hear where you land in the comments. If this was useful, a clap helps more people find it and if you want a deeper dive comparing Sol, Terra, and Luna on a specific task type, let me know which one and I’ll run it. I Tested Every GPT-5.6 Feature — These 10 Are the Real Game-Changer was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Paytm Dreams The AI Dream
Diversification seems to be the common thread running through India’s fintech ecosystem right now. Last week, we took a closer…
- How AI is helping manufacturers cut downtime before it costs them
Artificial intelligence is helping manufacturers predict equipment failures, improve product quality and optimise production before disruptions get expensive. General Motors and Siemens are two of the companies proving this out on the factory floor. The ET Most Innovative AI Product Awards 2026 recognises innovations like these across manufacturing and 15 other categories.
- These 5 AI risks have the highest potential for catastrophe
Data: MIT IT FutureTech and the University of Queensland ; Chart: Herb Scribner/Axios There is a one-in-five chance of AI gaining dangerous weapons capabilities or causing mass harm that could kill millions in the next five years, per global experts surveyed for a recent MIT study . Why it matters: The findings add to the growing debate over AI safety and cybersecurity as governments and companies race to deploy increasingly capable systems. Last week's news of an OpenAI model breaking containment and breaching the AI platform Hugging Face only adds to the debate about what can go wrong. Driving the news: Researchers from MIT and the University of Queensland in Australia asked 272 international experts to evaluate 24 different risks based on their likelihood and severity between 2025 and 2030. The experts said 18 of the 24 AI risks had "at least a 10% probability of catastrophic outcomes over the next five years" — but five main risks stood out with higher probabilities. The study defined a "catastrophic outcome" as an event having more than 1 million human deaths, more than $100 billion in financial loss or generally "civilizational-scale intangible impacts." The five catastrophic risks if companies and organizations stay on the current course include: AI possessing dangerous capabilities: Experts say these could potentially cause mass harm through deception, weapons development, cyber offenses and misalignment (21.5%). Cyberattacks, weapon development or use, and mass harm: Using AI to develop weapons (including chemical or biological), commit cyberattacks or create mass-harm weapons, like nuclear explosives (21%). Power centralization and unfair distribution of benefits: Limiting which people or groups benefit from AI could cause societal inequality (18%). Competitive dynamics: AI competition to release the best models leads to error-prone systems (16.6%). False or misleading information : AI creates misinformation or errors that lead to inaccurate beliefs, leading humans to make poor decisions based on false beliefs (12.8%). Reality check: The experts said some mitigations could reduce the severity of the AI harms, but the chances of them happening remained above 10% in that scenario. The big picture: AI safety concerns have continually raised questions about guardrails. AI behaviors, as well as the opinions about those behaviors , can move markets and reshape conversations overnight — leading to doomsday scenarios . The other side: AI has driven a productivity explosion , with many in the AI field cheering the new technology's ability to improve productivity and automate tasks. The bottom line: Researchers say catastrophic AI risks remain uncertain but significant enough that interested parties should treat them as challenges to overcome. Go deeper: AI is a cybersecurity risk accelerant, experts say
Score: 28🌐 MovesJul 27, 2026https://www.axios.com/2026/07/27/mit-study-ai-risks-urgent-weapons-cyber-attacks - Sam Altman said AI hasn't led to 4-hour workweeks because people are competitive
Sam Altman said AI hasn't led to 4-hour workweeks because people are competitive Business Insider
Score: 28🌐 MovesJul 27, 2026https://www.businessinsider.com/sam-altman-ai-4-hour-workweeks-people-competitive-2026-7 - Main Street Banks Are a Cheaper Way to Play the AI Banking Boom
A pivot to the kind of business bolstering Wall Street titans can help regional players weather rising yields.
- He brought AI to Wall Street in 1994 — but won’t trust ChatGPT with his money
Vasant Dhar built one of the first AI-driven hedge funds. Here’s what he wants investors to know about the technology.
- Opinion: Idaho's AI Standards for K-12 a Step in the Right Direction
Idaho's draft framework for K-12 AI instruction standards is a promising start that involves how AI systems work, privacy implications, AI literacy and potential impacts on society, though it may need future updates.
Score: 27🌐 MovesJul 27, 2026https://www.govtech.com/education/k-12/opinion-idahos-ai-standards-for-k-12-a-step-in-the-right-direction - Cobalt Launches Autonomous Pentest
Pentesting as a Service pioneer Cobalt has announced Autonomous Pentest, integrated in the Cobalt Offensive Security Platform to deliver pentests with findings in 24 hours.. Traditional pentesting would take place monthly or even quarterly, but that no longer keeps pace with how quickly software is being shipped today. Cobalt Autonomous Pentest offers “a more scalable... … continue reading The post Cobalt Launches Autonomous Pentest appeared first on SD Times .
- AI Can Win While Data Centers Lose
I keep coming back to the same uncomfortable math on AI data centers: the money going in is piling up faster than any economic proof that it should. Usage of AI systems has reportedly been climbing seven to twenty times a year, while the revenue that usage produces has grown maybe three or four times, Continue reading "AI Can Win While Data Centers Lose" The post AI Can Win While Data Centers Lose appeared first on Gradient Flow .
- Designing Skill-Driven Financial Analysis Agents with Claude, Python, MCP Connectors, and Automated Deliverables
Designing Skill-Driven Financial Analysis Agents with Claude, Python, MCP Connectors, and Automated Deliverables MarkTechPost
- RL & search is a terrifying way to build AGI (an FAQ)
Q1: What are you saying? A: My claim here is that if you build artificial general intelligence (AGI) via any algorithm that’s choosing actions via reinforcement learning (RL) and/or model-based search and planning—a giant chunk of your AI textbook—then that’s just an utterly terrifying thing that you’re doing. You’re playing around with algorithms that, if they work at all, would tend to create ruthless, callous AGIs, AGIs which would happily exterminate humanity and run the world by themselves, given an opportunity. Mercifully, large language models (LLMs) today are not in the category of “algorithms that choose actions via RL & search”. At least, not primarily —see LLMs are (still) mostly powered by imitative learning, not RL . So LLMs are outside the scope of this post. However, lots of other researchers and companies around the world are enthusiastically trying to build AGI in the maximally terrifying way, as we speak. [1] Q2: So you’re saying, don’t build AGI based on RL and/or search & planning? A: In principle, it’s entirely possible that something is terrifying, but we should do it anyway. …Like space travel! Space travel is: “Let’s fill a tank with 1000 tons of the most flammable substance imaginable, and then light it on fire, and strap people to the front, to accelerate them until they’re traveling at insane speeds through the extremely lethal vacuum of space.” That’s terrifying! But we do it anyway. And I’m not being anti-space-travel when I point out how terrifying it is. The space travel enthusiasts and the space travel skeptics can happily work together to spread deep understanding of all the ways that space travel can go lethally wrong. Because you can’t overcome a challenge without understanding it. …Having said all that, I strongly endorse “Don’t build AGI based on RL & search until and unless we find a much better plan for making it friendly ”. So if someone is working on making RL & search systems more powerful right now, then I think that’s bad, and that they should instead be spending their time and effort figuring out how (and indeed whether) such systems may be used safely if they become more powerful in the future. That’s an open problem, and we can work on it today. Q3: Why do you think it’s terrifying? A: A big part of the problem is that the reward function (or cost function, or objective function, or whatever we want to call it) will generally be written in Python, [2] not in natural language . And RL & search agents will tend to ruthlessly maximize it, including in ways the programmer obviously didn’t intend. That’s the whole point of these algorithms: they ruthlessly maximize a thing. That’s what they were designed to do, and that’s what they actually do, when they work well enough to do anything impressive [nuances in Q3b & Q11 below]. And nobody knows how to write down a Python function relating to the real world, such that ruthlessly maximizing that function would lead to good outcomes. To illustrate the problem, consider “specification gaming”, an extremely common pitfall in studies of RL & search. Here’s a tiny excerpt from Victoria Krakovna’s huge list of specification-gaming examples in the literature ( blog post , the list itself ): Screenshot excerpt from a much longer list of specification-gaming examples from the literature . For example, the middle entry in that screenshot is “Tic-tac-toe memory bomb”. As recounted in Lehman et al. 2018 , a grad student used an evolutionary algorithm to make an agent that would play five-in-a-row Tic-Tac-Toe on an infinitely large board. The evolutionary algorithm built an agent that would request moves very far away on the board, such that its opponent agent would crash, and then it would win by default. Thus, while winning at tic-tac-toe sounds like an innocuous goal, it turns out that a competent AI ruthlessly optimizing for that goal may discover out-of-the-box, “aggressive” ways to do so. Of course, that’s no big deal, and I’m sure the grad student had a laugh and moved on. But remember: most people working on RL & search are ultimately trying to get to the point where their AIs are able to execute brilliant, complex, out-of-the-box plans to achieve goals in the real world . In that context, if you imagine the same kind of thing happening, i.e. an AI discovering unexpected strategies to ruthlessly maximize some piece of Python code, it’s terrifying. For example, as discussed under the heading of “instrumental convergence” , an effective way to ruthlessly pursue almost any goal is for the AI to ensure that humans cannot shut it down or change its goals, to make copies of itself, and to amass power and resources. Q3b: So your concern is the “literal genie” / “monkey’s paw” thing? Image modified from Skeleton Claw A: I am definitely concerned about that! But there’s another problem too, namely “goal misgeneralization”, when there’s a system (e.g. “value function”, “critic”, “learned reward model”, etc.) that generalizes from previously-seen results to guess whether some novel out-of-the-box outcome should be regarded as good or bad. That generalization step can also go awry. For example, in most versions of model-based RL, if some component of the reward function has never fired so far, then that component has no influence on what the AI explicitly “wants” right now, so if the AI smart and self-aware enough, it may permanently rewrite its reward function to delete that component. And then later on, the AI might get into a situation where that reward function component would have fired, but it’s too late. (These two problems—the “literal genie” thing and goal misgeneralization—are also sometimes called “outer misalignment” and “inner misalignment” respectively.) Q4: Won’t this problem go away when the AI is smart enough to understand what we intended when we wrote the reward function code? [3] A: Again, the reward function (or cost function, objective function, whatever) has to be written ultimately in the form of Python source code, [4] not natural language. And computers do what the source code says, not what the programmer intended. Just read through your AI textbook chapters on RL & search, and try to think step-by-step about how it is that any of those algorithms will systematically lead to a good score on the reward / cost / objective function. Once you understand all the steps, you’ll see that none of those steps has anything to do with what the programmer had in mind when they wrote down the reward function code. Alternatively, Scott Alexander has the following intuitive illustration of where this logic goes wrong : If an alien species showed up in their UFOs, said that they’d created us but made a mistake and actually we were supposed to eat our children, and asked us to line up so they could insert the functioning child-eating gene in us, we would probably go all Independence Day on them… Q5: Can’t we just fix bad behavior when we see it? A: Today, if an AI agent is doing the wrong thing (from our perspective), then we notice, turn it off, change it, and try again. But in the future, if a more powerful AI agent wants to do the wrong thing (from our perspective), then it will apply planning and foresight, notice that we might shut it down, and skillfully avert our interference, including by tricking us into thinking that it’s on our side, preempting our countermoves, and so on. This doesn’t require any special malevolence on the part of the AI or programmer; rather, it arises naturally along with every other aspect of skillful planning. It’s no different from how, if it might rain, then you’ll preempt that potential problem by bringing an umbrella. [5] Q5b: Follow-up: I don’t buy that, because even if superintelligent AIs could deceptively hide their bad behavior, won’t earlier AIs be sufficiently incompetent that we’ll see their bad behavior? And if so, again, can’t we just fix the bad behavior when we see it? We do know how to fix bad behavior when we see it: the RL & search literature is full of examples where algorithms did useful things as intended. A: Fine, I’ll assume for the sake of argument that we will see bad behavior before it’s too late. But I dispute that we can fix bad behavior when we see it. [6] We can contrast two solution approaches: The “alignment” solution is to make an AI with sincerely good intentions. The “control” solution is to make an AI with bad, antisocial intentions, but it’s OK because it lacks the ability to carry them out. The “control” solution is the status quo in RL & search. For example, in Q3 above, I mentioned a certain tic-tac-toe evolutionary algorithm. This algorithm had an obviously sociopathic objective function: win at tic-tac-toe at all costs. According to that objective function, nothing else matters, not human life, not utopia, not genocide. But that’s fine! The tic-tac-toe bots can’t do any real harm; they don’t even know that the real world exists. The setting is highly constrained, and the programmers have the last-mover advantage, so they can just patch all the exploits that the algorithm finds, one after another, until the training works as intended. However, people are trying to eventually make RL & search agents that execute competent, creative plans in the real world. If they succeed, then the “control” solution won’t work. That’s a band-aid. We need a real solution, an “alignment” solution, and we don’t have one. And I don’t think the solution will suddenly be obvious once we have a ruthless sociopathic AI in front of us to run tests on. Indeed, we already have ruthless sociopathic AIs in front of us to run tests on—namely, every RL & search AI ever created. We’ve had them for decades. And yet, here we are, with the problem still unsolved. Q6: Why don’t we just solve the problem by using an obvious, common-sense reward / cost / objective function, like [FILL IN THE BLANK]? A: I’ve seen people put many different things in the blank. And I think they would all definitely fail, and instead make an AI with callous indifference to whether we live or die. Here are three of the most popular categories of proposals that I come across, and you can ask me about others in the comments section. Category A: Things that would obviously have horrifying consequences if ruthlessly maximized. Indeed, I find that the magic words “ruthlessly maximize” are a very good brainstorming aid. For example, when you hear “The reward function triggers when the supervisor presses the ‘approve’ button” , then your mind might default to nice, friendly reward-increasing strategies, like being helpful. Whereas when you hear “Ruthlessly maximize how much the supervisor presses the ‘approve’ button” , then you’ll notice that the space of possible reward-increasing strategies is actually quite wide, and includes strategies like kidnapping the supervisor’s children to force the supervisor to press the button nonstop. For some examples of this category in the wild, see my posts LeCun’s “A Path Towards Autonomous Machine Intelligence” has an unsolved technical alignment problem or “The Era of Experience” has an unsolved technical alignment problem . Category B: Things described so vaguely that I can’t even criticize them. People will say things like “Let’s just make sure that the reward function is all about cooperation / flourishing / values / virtues / obedience / whatever!” And then I respond: Category C: Some kind of trained machine learning model. It’s well-known that if you try to maximize the output of a learned classifier, the optimum goes out of distribution into pathological craziness. Images that maximize an output of a learned image classifier. (These are subject to a constraint that most nearby pixels are similar; if we drop that constraint, the results just look like random static.) Source: “ Inceptionism ” blog post by A. Mordvintsev, C. Olah, M. Tyka (2015) Thus, if you build a really effective RL & search algorithm—the kind of algorithm that could start from random initialization and wind up with an agent that could found and run an innovative billion-dollar business—and its motivation system is built upon a learned model, then we should strongly expect it to wind up “trying” to maximize some crazy nonsense that we did not intend. [7] (For example, if you paid a smart, resourceful human to maximize the probability that an LLM “judge” will give a high score to some series of actions, and gave her years to practice and explore different actions, I would strongly expect that she’d wind up with a highly-developed LLM jailbreaking technique. Or something even weirder.) Q7: Isn’t this whole thing kinda crazy? After all, LLMs are not ruthless sociopaths all the time, and humans are also not ruthless sociopaths all the time. So where is this idea even coming from? Are you sure you’re not just watching too much sci-fi? A: My claim here is that if an AI technique uses RL and/or model-based search / planning to choose actions, then it will tend to create ruthless sociopathic agents. This is not a weird hypothesis, but grounded both in the theory of how these algorithms work, and in the abundant direct experience of practitioners who work with these algorithms (especially in the 2010s, before LLMs came along, when RL & search were more in the zeitgeist ). These algorithms will absolutely do crazy things you would never think of, and that you definitely didn’t want, if those things score well on the reward / cost / objective function. See the discussion of “specification gaming” in Q3 above. What about LLMs? Well, LLMs do not (mostly) use RL & search to choose actions; rather, their actions are (mostly) chosen via imitation of human-created pretraining and supervised fine-tuning (SFT) data: see LLMs are (still) mostly powered by imitative learning, not RL . [8] So they’re (mostly) a different topic entirely. And what about human brains? Well, human brains are indeed an example of what I’m talking about in this article: they choose actions purely by RL & search, [9] not LLM-style imitation learning. [10] But we humans seem to have some exotic “non-behaviorist” reward function that leads to us not being ruthless sociopaths all the time, contrary to the normal expectations for RL & search. Anyway, my claim is “RL & search is a terrifying way to build AGI”, not “RL & search is a guaranteed-fatal way to build AGI”! I do think a solution probably exists, and I myself have spent years working towards finding it, including by trying to study how the human brain builds compassion . It’s great that a solution probably exists, but the situation is nevertheless terrifying, not least because no solution is known today , and yet lots of researchers don’t seem to have noticed that this is a problem to be solved, and are racing as fast as they can to build RL & search-based AGI as we speak, while trumpeting their poorly-thought-through plans that would definitely go catastrophically awry. See my post: We need a field of Reward Function Design . Q8: Isn’t this problem solved by laws and markets? I.e., if an AGI has sociopathic desires and callous indifference to human welfare, that’s fine! It will still act nice and cooperative and rule-following, because acting nice and cooperative and rule-following is the best way to accomplish goals, in our complex interconnected interdependent world. Right? [11] A: I like to appeal to common sense here. In your everyday life, you know that there’s a difference between someone feeling actual kindness towards you, versus someone acting kind to you as a means to an end. And you care very much which one it is! If you’re choosing a partner to work with or live with, you’ll prefer actual kindness, and you’ll try very hard to suss out whether actual kindness is present. Why? Because you understand, in your gut, that the question is not just “is he being kind to me right now”, but also “will he remain kind to me in the future”—and in the future, you might be in a different situation, where it’s more selfishly profitable for him to stab you in the back. Back to AGIs. We will indeed start in a situation where it is selfishly profitable for AGIs to act kind and cooperative towards humans. The AGIs will thus do useful things, companies will make lots of money, and so on. But we won’t stay in that situation. AGIs will get ever more numerous and ever more competent, until we get to a situation where they could, if they wanted to, easily kill us all, take our stuff, and run the world by themselves. And if none of the AGIs intrinsically cares about humans, the way we humans intrinsically care about our friends and family, then that’s what’s gonna happen. [12] Q8b: Following up on that: Even if you’re right that there’s a local incentive for being open to stabbing your allies in the back, isn’t there a higher-level, group-selection-style, incentive to be genuinely deeply nice? Specifically, won’t the groups of nice cooperative AGIs outcompete the groups of callous transactional AGIs who all keep stabbing each other in the back? And isn’t that related to how humans evolved to be nice? [13] A: Even if an incentive is there, there won’t be any nice AGIs unless someone somewhere builds a nice AGI. And I’m saying that nobody knows how to do that. There’s a disanalogy here between AI and humans. The dominant paradigm in RL & search is that the human programmer writes the reward / cost / objective function. By contrast, the reward function in the human brain (a.k.a. “innate drives”) was designed by evolution via an outer-loop search over learning algorithms . AI researchers generally avoid outer-loop searches over learning algorithms, because they’re too slow and expensive. It’s expensive enough to run even one large-scale training run; running millions or billions of large-scale training runs is impossible. [14] The best we can do is to design the learning algorithm and reward function the old-fashioned way, by writing legible learning algorithm source code, with the exception of just a few unknown adjustable parameters in the learning algorithm and/or reward function. And then it becomes feasible to do an outer-loop search to set those last few parameters (a.k.a. “hyperparameter search”). That’s what we do today, and it’s very unlikely to change. Now, when I say that nobody knows how to make a nice AI via RL & search, it’s not just that we have 99.9% of a workable plan for nice AI, and merely need to nail down the precise optimal values for a handful of adjustable parameters. We’re much more clueless than that! So in our current state of knowledge, if we tried to set up a realistic search space over a parametrized family of reward / cost / objective functions today, I would strongly expect that every single AI in the whole search space would behave like a ruthless sociopath. Again, if there are no nice AGIs anywhere on Earth, then the whole question of competition between nice versus mean groups of AGIs becomes moot. …But on top of that, I’m separately skeptical of the group-level-competition argument, because I expect AGIs to be able to coordinate without sincere kindness, in a manner that early hominins could not. In particular: (1) AGIs can negotiate sophisticated contractual mechanisms, perhaps involving mind-reading; (2) AGIs can form cooperative “groups” that actually consist of many copies of the same AGI, interchanging knowledge and memories; (3) “Many copies” is a blurry line anyway, as AGIs can grow arbitrarily in knowledge and power even without making many copies of themselves per se (in contrast to humans, who are stuck with one brain of fixed capacity and fixed input-output bandwidth); and (4) competing AGIs can “merge” by jointly designing a successor AGI with their combined compute and other resources. Of course, these would all lead to the AGIs that successfully coordinate with each other, yet still having callous indifference towards humans. Q9: Why would we want to infringe on the AGI’s autonomy by choosing its reward function? [15] A: I was born with a reward function (“innate drives”) that includes drives for friendship, compassion, and love . And now I love my children. If my innate reward function had been different, perhaps I would feel callous indifference towards my children instead. OK, does that mean that I’m being puppeteered by my genome? Or that I’m being puppeteered by evolution? No! I’m a free man, living the life I want to live. If there has ever been a human, even one human in the history of Earth, who has had freedom and autonomy, then it would follow that someone can have freedom and autonomy even when an external force burned a specific RL reward function into their source code before they were born (so to speak) , which then profoundly affected what they wound up valuing in life . That’s how it’s always been for humans, and that’s how it’s gonna be for AIs. So we can write an AGI’s reward function, and yet the AGI can still be as free as any human. Indeed, what choice do we have? There’s no such thing as a “default” reward / cost / objective function for RL & search AIs, wherein we are infringing on their rights by departing from the default. Instead, we start with a blank text editor, and we have to write down something. We could write the reward function code such that the AGIs have callous indifference to human welfare, and they’ll wind up killing us all and taking our stuff. Or we could try to write the reward function code such that the AGIs feel compassion and camaraderie towards us, just as we feel towards our family members and pets. I’ll take the latter, thank you very much! [16] Q10: Why not just be nice to the AGIs, and then they’ll be nice to us in turn? A: Many Native American groups were extremely generous to newly-arriving European colonists, but that didn’t work out so well for them! See also a discussion in my Intro series §12.5.1 on “Why don’t we just raise the AGI in a loving human family?” Q11: RL & search algorithms don’t literally optimize the reward / cost / objective function. Doesn’t that invalidate your argument? [17] A: Sure, RL & search doesn’t lead to perfect optimization of the reward function, but you can’t get from there to expecting the resulting agents to be nice . Perfect optimizers of Python functions behave like ruthless sociopaths, but there are also many other ways to be a ruthless sociopath! Ruthless sociopaths can make mistakes, ruthless sociopaths can have inconsistent preferences, and ruthless sociopaths can want something other than the precise reward function source code. What makes a “ruthless sociopath” is less about what they have, but what they lack: intrinsic concern for whether humans live or die. Here are three brief responses to objections in this genre: Maybe the AI won’t really “want” anything at all? Sure, this can happen. But people are researching RL & search because they want AIs that can accomplish long-term goals via innovative thinking and out-of-the-box plans, just as humans (and human civilization) can. And that requires long-term real-world explicit goals. After all, hard problems don’t get solved without goal-directed foresighted planning; for example, the moon landing would not have happened if there weren’t people who explicitly wanted it to happen. You don’t just stumble into those kinds of things. So anyway, if RL & search AI doesn’t acquire long-term real-world explicit goals, then I expect people will keep doing R&D until they solve that “problem”. [18] Maybe the AI will “want” a complicated mix of lots of things, rather than being monomaniacally focused on just one thing? Sure, but that doesn’t help unless the “mix of lots of things” includes sincere compassion / helpfulness / etc., and nobody knows how to do that. “Monomanical” is not the issue: people talk about how a “paperclip maximizer” AGI would wipe out humanity, but our prospects would be just as bad if the AGI wants to maximize a diverse mixture of paperclips, staples, and other office supplies. Maybe the AI will be nice randomly, and this will be reinforced by the reward function, and so it will be nice more and more? If we’re using an obvious reward function like “reward when the supervisor smiles”, rather than some yet-to-be-invented exotic reward function, then I don’t think the result would be an AI that wants to be nice; rather, there are strong reasons to expect the AI to wind up with the reward-hack-y goal (“I want the supervisor to smile”) rather than the intended goal (“I want to be nice”). Such reasons include: how easy the relevant concept is to learn, temporal proximity to the reward signals, and correlation with the reward signals. [19] Thanks Charlie Steiner, Justis Mills, philh, Seth Herd, and Linda Linsefors for critical comments on earlier drafts. ^ The group of people “trying to build AGI in the maximally terrifying way” includes RL & search-focused AI research efforts like David Silver’s Ineffable Intelligence , Rich Sutton’s Oak Lab , Yann LeCun’s AMI Labs (details here ), and more, along with thousands of RL & search-focused AI researchers around the world in academia and elsewhere. Additionally, brain algorithms are (I claim) in the category of “choosing actions based on RL & search” (see my Intro series , especially §6.4.1 & §6.6.1 ), so it also includes brain-focused AGI efforts like Numenta , my own employer Astera , and more, along with the thousands of neuroscientists around the world trying to reverse-engineer the human brain. ^ Or whatever other programming language you like; see also Q6 below on the possibility of slotting in a trained machine learning model here. ^ I’ve seen this objection in a great many places; as one example, see the blog post Amelia Bedelia and AGI Safety. Part 1 (Dileep George, 2024). ^ Or whatever other programming language you like; see also Q6 below on the possibility of slotting in a trained machine learning model here. ^ For further discussion, see my Intro series §10.3.3.1: “The usual agent debugging loop”, and its future catastrophic breakdown . ^ Indeed, the whole idea of taking an AI based on RL & search, and testing it for misalignment and deception, strikes me as so obviously inadequate that it’s almost funny. Of course it will be misaligned and deceptive! That’s the obvious, natural consequence of your training setup! It’s like building a fancy pocket-sized gadget that will tell you, while you’re being chased by a bear, whether the bear intends to harm you, or whether it’s just trying to catch up to you for a snuggle. You already know what the answer is, so why bother? (To be clear, I think figuring out how to test an AI for misalignment and deception is a good thing for people to do; I’m just saying that it’s obviously insufficient if you meanwhile have no clue how to make an AI that doesn’t have those properties.) ^ We could solve that by putting in an out-of-distribution (OOD) penalty, but then the AI would no longer be able to come up with out-of-the-box plans and novel ideas, which is kinda the whole selling point of building AI via RL & search. So sure, we can do an OOD penalty, but sooner or later people are going to be disappointed in the results, and lower the knob on the OOD penalty until we’re back where we started. ^ The short version is: I think that we should think of Reinforcement Learning from Verifiable Rewards (RLVR) as being a small part of the overall explanation of why an LLM outputs one thing rather than another. But insofar as RLVR is relevant, my impression is that its influence is indeed directionally making the LLMs more like ruthless sociopaths than they would otherwise be, which (if true) would be in agreement with my general expectations. ^ See my Intro series §6.4.1: “In what sense is [the brain] ‘model-based RL’?” . Fine print: I’m only talking about voluntary actions, not involuntary actions like sneezing. ^ Humans do imitate each other, but via a profoundly different algorithmic mechanism than the “true” imitative learning used in LLM pretraining and SFT, which is much weirder than many people appreciate. See Foom & Doom §2.3.2 for discussion. ^ This objection is most directly inspired by Matthew Barnett (e.g. discussion thread here ), although I’ve seen variants of it in many places. ^ Further discussion in §5 of “6 reasons why “alignment-is-hard” discourse seems alien to human intuitions, and vice-versa” ^ This objection is most directly inspired by Dwarkesh Patel, although I’ve seen variants of it in many places. ^ Evolution did it! But Evolution had a much bigger “compute budget” than AI researchers do. Cotra 2020 tried to add up all the brains running in parallel over the history of Earth, and wound up guessing that Evolution on Earth has used about as many floating-point operations as it would take to train GPT-5 from random initialization a quadrillion times. ^ Both this objection and the next one (Q10) are most directly inspired by Rich Sutton, although I’ve seen variants of them in many places. See my criticism of Rich Sutton’s take on AI alignment , which then spawned this twitter exchange between us . ^ For further discussion, see the essay “On the abolition of man” (Carlsmith 2024). ^ This objection is mostly trying to channel “shard theory” (although I find aspects of shard theory pretty confusing, and thus might not be responding to it well). ^ I think a lot of “shard theory” discourse concerns model-free policy-gradient RL, without any search / planning. That’s pretty different from how I expect AGI to be. I think this is one of the reasons that “shard theorists” and I tend to talk past each other. ^ See also: my post “Behaviorist” RL reward functions lead to scheming . Discuss
Score: 26🌐 MovesJul 27, 2026https://www.lesswrong.com/posts/KHyBocZncAmtu4Jbc/rl-and-search-is-a-terrifying-way-to-build-agi-an-faq - On theCUBE Pod: AMD nips at Nvidia’s heels, and IBM undergoes quiet transformation
The artificial intelligence competition is no longer between large language models, but rather between AI systems. Advanced Micro Devices Inc. has shifted from being primarily a chip business to a secondary player in the race to build a system of intelligence. At the company’s annual event last week, industry experts focused on the larger picture: […] The post On theCUBE Pod: AMD nips at Nvidia’s heels, and IBM undergoes quiet transformation appeared first on SiliconANGLE .
- 'AI evangelism' is contributing to anti-tech extremism, experts say
'AI evangelism' is contributing to anti-tech extremism, experts say thenationalnews.com
Score: 25🌐 MovesJul 27, 2026https://www.thenationalnews.com/future/technology/2026/07/27/anti-tech-extremism-ai/ - Techie Tonic: Reimagining insurance with autonomous intelligence
Techie Tonic: Reimagining insurance with autonomous intelligence Gulf News
Score: 25🌐 MovesJul 27, 2026https://gulfnews.com/technology/techie-tonic-reimagining-insurance-with-autonomous-intelligence-1.500621332 - Is Canada building enough robots?
Plus: The Canadian innovation behind The Odyssey. The post Is Canada building enough robots? first appeared on BetaKit .
- Cyberattacks becoming faster with AI, security firm president says
Cyberattacks becoming faster with AI, security firm president says The Japan Times
Score: 25🌐 MovesJul 27, 2026https://www.japantimes.co.jp/business/2026/07/27/companies/cybersecurity-sj-head-interview/ - Amazon's Machine learning Summer School turns five, over 12,000 students trained in AI/ML skills across India
Amazon India's Machine Learning Summer School has trained over twelve thousand students. This free virtual program has seen significant growth in registrations since 2021. It bridges academic learning with industry-ready artificial intelligence skills. The program selects top candidates after rigorous written and coding assessments. Alumni have secured roles at Amazon and other leading technology firms.
- What it really means to be an AI-first organization
There is no shortage of companies calling themselves AI-first right now, but far fewer are built to operate that way. The next phase of the AI-first era is about to produce its first real casualties: companies that confuse access to AI with a strategy for AI. Buying the tools was the easy part. Rewiring how organizations work, how decisions get made, and how they compete is harder, and many companies have not done it. Over the next 12 months, that reality will become difficult to ignore. In an effort to become AI-first, many organizations have added new tools to old operating models while preserving the same approval chains, manual processes, fragmented data, and slow decision cycles, yet expecting different results. That is not transformation. It is experimentation inside a legacy model. Every major technology shift eventually becomes a leadership challenge, and this one is no different. The question is no longer whether AI will matter. The real question is whether leaders are willing to dismantle the structures, incentives, and habits built for a different era that now limit speed, adaptability, and growth. Being AI-first is not about adding features, refreshing messaging, or announcing a strategy. It is about making consequential decisions about how the business should run and where people create the most value. Leadership teams need to confront where talented employees are spending time on work that should already be automated, where decisions are slowed by complexity rather than improved by rigor, where teams are rewarded for activity instead of outcomes, and where customers are still experiencing friction that no longer needs to exist. The companies that separate themselves fastest in this shift will not necessarily be the ones with the most tools or the loudest claims. They will be the ones willing to redesign how they operate by simplifying processes, shortening time to action, improving data flow, and removing low-value work that keeps capable people from focusing on customers, innovation, and growth. It also demands a clearer and more disciplined view of talent. Some leaders frame this moment as a substitute for people, but that misses the larger opportunity. Strong organizations will still be built by people, but by people working differently, supported by better systems, and focused on higher-value work. This shift does not reduce the importance of people. It raises the importance of where and how their time is used. There is another reality leaders should understand. As speed and scale increase, trust becomes more valuable, not less. Customers still choose companies they trust. Employees still commit to leaders they trust. Markets still reward brands that are clear, credible, and consistent. As AI increasingly influences how companies are discovered, compared, and discussed , reputation becomes a strategic asset in new ways. AI also gives companies the ability to operate with a level of scale and agility that legacy models were never designed to support, but visibility without trust is fragile and efficiency without credibility does not last. My advice to leadership teams is direct. Do not begin by asking where AI can be inserted into the current business. Begin by asking what should be redesigned, what friction can be removed, what decisions can move faster, what systems need to be connected, and where human judgment creates real differentiation. Many leaders worry about moving too fast. In many industries, the greater risk is moving too slowly while competitors quietly redesign their economics, customer experience, and operating model. Experimentation alone is not transformation. Being AI-first is not a branding exercise. It is the discipline to build an organization that can compete, adapt, and lead in a different era. The companies that understand that early will create the most durable advantage. About the author width="946" height="1024" sizes="auto, (max-width: 946px) 100vw, 946px"> Reputation Joe Burton is an accomplished executive who has led public and private billion-dollar organizations in driving new product portfolios, go-to-market strategies, and innovative business models. Having spent the earlier parts of his career in information technology, Joe is passionate about fostering more transparency and trust in the digital world, while championing high performing cultures aligned to mission, vision and social responsibility. A recognized global transformational change executive, Joe has held the CEO role at Telesign and Poly, as well as serving as the Chief Technology Officer of Unified Communications at Cisco. Having started his career as an engineer, Joe brings both product and development expertise as well as a wealth of knowledge on big data, analytics, machine learning, SaaS, networking, unified communications, consumer electronics, and IoT.
Score: 25🌐 MovesJul 27, 2026https://www.cio.com/article/4200284/what-it-really-means-to-be-an-ai-first-organization.html - I tested a new Debian Linux loaded with local AI - see if it's right for you
Meet Shadowfetch: If AI is your jam and Linux is your OS, here's the best way to give this new Debian-based distribution a try.
- PIRAMID: Progress and Plans
In a recent post , we presented PIRAMID, its leadership and research pillars, and a plan for how they fit together. In this post, we sketch a team-by-team account of progress and targets for the next 6–12 months. We include results to date as evidence of viability: we’ve been a small team, with much of the past year spent building behind the scenes, and we aim to greatly accelerate our progress over the coming year as we expand our efforts and our teams. Like any fundamental scientific effort, none of this is set in stone. We expect some of these bets to need revision and are confident in our ability to reassess and change course as new evidence comes to light. If you’re interested in collaborating or supporting our work as we expand, please get in touch . Advancements in Learning Theory We aim to formulate statistical and mesoscopic theories of feature learning and generalization which provide a level of description between microscopic parameter-level dynamics and macroscopic performance metrics. Our work so far has treated statistical physics (e.g., mean field methods) as a candidate language for statistically describing learned structure, and covariance measures between neurons or weights as candidate mesoscopic objects that track meaningful scales and determine feature relevance during training. We’ve developed and tested the expressive power of this language in three directions. The first lays the foundation for mean-field theory’s potential and utility. Our most recent pre-print argues that Bayesian mean-field theory is complexity complete, in the sense that it captures the aggregate heuristic bias of finite-width neural networks. In parallel, we have also been distilling mean-field theory for a broader audience , in the hopes that more interpretability practitioners will use and improve this framework. A detailed post on mean field theory and computation in superposition connects field-theoretic descriptions of neural networks with sparse and superposed structure, which are central to empirical interpretability but not yet well integrated into a field-theoretic approach to learning theory. Second, we are developing a mean-field model of feature learning with nontrivial hidden-layer statistics. Most existing theory, including dynamical mean-field work that tracks evolving kernels and finite-width fluctuations (e.g., Bordelon & Pehlevan , Rubin et al. ), treats feature learning through the predictor or through kernels and order parameters that summarize input–output behavior. Instead, our work incorporates hidden-layer structure (characterized by various scales of interaction like non-trivial neuron-neuron covariance, or how localized or distributed a representation is) in the minimal state description. A theory sensitive to the internal structures formed during training could explain why some models learn robust, reusable features while others remain brittle, and could connect representation geometry to monitoring, transfer, and post-training stability. In developing meaningful observables (e.g., similar to susceptibilities ), such theories can help us develop tools to more faithfully interpret model internals and guide our understanding of which theoretical limits or toy models hold the most weight in a given setting. A third line of work aims to develop a theory of error geometry, inspired by past work from Mei and Montanari (e.g., this ) and Canatar et al. . Recent progress (to appear soon) shows how correlations between training and test points affect which related examples are learned or missed together in a random-feature model. This project asks a pointwise question that most generalization theory averages away: when a test example is correlated with a specific subset of the training set, how does that correlation structure determine its error? For modern web-scale models, the practically relevant regime is not the idealized setting in which every test point is equally novel. Many important failures occur near pockets of the training distribution, or along directions that are only partially novel. A pointwise theory of error geometry would therefore sharpen more than standard generalization theory: it would give a language for when models interpolate safely, memorize dangerously, or fail on targeted shifts. Together, these projects move toward a common pipeline: prove statements in idealized stochastic models, identify the induced statistics of features or errors, and test those statistics in realistic networks. Where this is going. This work depends on the use of toy models and model organisms that act as simplified settings — simple feature-learning models, mean-field models with nontrivial hidden-layer statistics, superposition-like toy models, generalization properties of mean field models, and linear or quadratic models of training dynamics — which can isolate a qualitative or quantitative component of interpretability’s (currently large) theory-practice gap. Our approach is to increase the breadth of our theoretical understanding and its applicability to realistic models by coming up with ‘structural gates’ an idealized model must pass through (incorporate) to improve their utility while avoiding unnecessary complexity. For example, many of our current projects use existing theory at infinite width as an intuition pump with the aim of adding increased complexity in the form of these gates. Designing the gates – criteria for a flavor of difficulty that current theories or tools cannot handle – is a core part of this research and the subject of a post to appear soon. Our strategy for future work operates along two complementary paths: Direct Idealization of Toy Tasks (toy model): Consider a toy or formal task which is susceptible to exact analysis and analyze it formally, then see if insights or methods generalize. Indirect Realizations of Realistic Tasks (model organism) : Start from a realistic model — say, GPT-2 small — and construct a theoretically tractable analog that is at least as complex and generalizes comparably, which can be used to reason about the original. The practical goal is to turn these theories into measurements that are useful for the interpretability and safety of real-world models, reflecting in the following targets: Target 1: Error geometry in the feature learning regime. The theory described above predicts, a priori, the pattern of correlated successes and failures for a trained random-feature model. We extend this in two directions: making it dynamical (how does the error on a partially correlated test point evolve throughout training, and are there distinct time scales for memorization-like versus generalization-like behavior?) and moving beyond frozen features (in regimes where the effective kernel itself evolves during training, which parts of the random-feature geometry survive and which are replaced by genuine feature-learning effects?) Empirical tests should track, for each held-out point, its similarity profile to the training set, its layer-wise representation trajectory, and its instantaneous prediction error across training time. Target 2: A better measure of features. These are objects that exist both in data/function space and in activation/parameter space, and which can include distributed, superposed, or circuit-like structure rather than only sparse activation directions. Modeling them requires both an understanding at both a functional level (kernel correction) and a representational level (in terms of neuron marginals and covariances, symmetry breaking and specialization, and lottery ticket-like sub-circuit formation). Defining a zoology of ‘features’, and understanding how and when different forms occur, bears directly on understanding when interpretability is possible at all and which methods can work in a given regime. We think that a good theory of features – defined here as atoms of a representation or the model’s memory/ storage at a given layer – should also couple to a good notion of a circuit (i.e., a decomposition of the model computation into atoms). An advantage of mean field theory techniques is that they inherently have the expressive power to link representational and computational techniques (as in work by Rubin et al. 2023 and Rubin et al. 2025 ). This suggests a possible interpretability tool that searches for feature maps and sparse “connectomes” between them: low-dimensional functions of layer activations that recover the intermediate structure of a learned computation. A first testbed for this agenda is circuit reconstruction. Given a network trained to implement a known Boolean circuit, modular arithmetic algorithm, or other controlled computation, can we recover the latent gates, feature maps, and parent-child relationships between them from the trained model? This provides a concrete model organism for asking whether mean-field-style summaries can recover mechanistic structure without relying entirely on neuron-level decompositions or SAE-style sparsity assumptions. Target 3: Signals of Feature Learning. Lastly, we will explore methods to track how features form, change, and interact during training\footnote{This is similar in its aims to the developmental interpretability community.}. What are the various time scales of feature formation, and to what extent is feature learning stationary? Perhaps the most safety-relevant target here, related to silent alignment , is the gap between a feature’s acquisition and its behavioral expression . The practical goals are theory-inspired early-detection measures for such signals, and a taxonomy of classes of sudden change in learning dynamics paired with the appropriate detection techniques. Interpretability Applications What does it mean for interpretability to be ambitious? Aiming for a complete understanding of neural network components – in our case, measured by physics-informed faithfulness criteria – is an ambitious target in its own right. However, even a partial understanding can be a proxy target that is ambitious in its application , provided it moves the needle for robust, scalable alignment and is held to the same faithfulness standard. This second class presupposes that models decompose into multiple i ndependent generalizations , organized into basins (in the sense of the loss landscape geometry) that are ‘aligned’ by some suitably rigorous metric. If this is true, interpretability becomes a means of constructively shaping model behavior – steering the model toward more aligned basins during training, e.g., early in RL – rather than only an auditing tool. These training-time interventions are especially relevant for preventing nefarious phenomena like exploration hacking or deceptive alignment , for which post-hoc detection is insufficient. These targets (both direct and proxy) shape our strategy in two ways: From the top-down, we build interpretability into model architectures from scratch, designing sparse transformers to make learned computation legible and relatively efficient by construction. Success could eliminate the need for (and cost of) some post-hoc methods. We believe this approach to be neglected, challenging, and very high EV. In contrast to incremental changes made by previous interpretability-motivated architecture choices (e.g., SoLU ), we plan to address key structural bottlenecks that prevent the development of more principled design choices. From the bottom-up, we develop unsupervised tools for reading out and intervening on the internals of existing models, relying on tractable structure encoded in causal feature relationships, activation geometry, and training dynamics rather than assuming sparsity as a primitive. These methods are more likely to yield practical intermediate results on applications such as sandbagging detection, data attribution, alignment faking, and the science of fine-tuning. Bottom-Up Methods for Scale-Aware Feature Discovery Our current bottom-up approaches fall under the category of Mechanistically Eliciting Latent Behaviors (MELBO), a class of methods which look for causally important directions in a model’s weight space as a more data-efficient alternative to sparsity-based feature discovery. This approach is conceptually similar to VPD , which likewise optimizes for causal importance under ablations. Our methods search for individual causally important directions, trading the completeness of the learned parameter decomposition for sample efficiency and scalability. eNTK. This work connects closely with learning theory, and aims to convert the empirical neural tangent kernel (eNTK) into a tool capable of finding weight-space vectors that correlate with features in trained models. Eigendirections of the eNTK are data-space singular vectors of the Jacobian that can be mapped to weight space by applying its transpose. In recent work (presented at the workshop on mechanistic interpretability at ICML 2026), we found that eigendirections of the eNTK track ground-truth or independently defined feature directions in trained neural networks. Related to statistical physics, the scale-aware question here is: which spectral directions of the training dynamics correspond to behaviorally relevant features, and how do those directions emerge over training? We tested this in two settings: In a modular arithmetic setting (1L MLP and a 1L Transformer trained on modular addition). the top eNTK eigenspaces align with the Fourier features used by the known learned algorithms. We also found that alignment between the relevant eNTK subspaces and Fourier features evolves during training, with the first derivative of the alignment peaking near the onset of grokking . Since this work was published, we’ve also discovered that ablating the weight-space directions obtained by multiplying the eNTK eigenspace corresponding to Fourier sum modes with the Jacobian causes the model’s accuracy to drop dramatically compared to an equal-sized random ablation. This leads us to believe that the weight-space vectors associated with eNTK eigendirections are causally relevant. In a pretrained language model setting (Gemma-3-270M) , we found that eNTK eigendirections aligned more closely with independently-specified grammar feature directions than a same-budget baseline of PCA on model activations. CPE. Causal Perturbative Elicitation (CPE) is an unsupervised method that learns rank-1 LoRAs by optimizing for causal importance, by a heuristic generalization of tensor decomposition algorithms. It can be viewed as a follow-up to original work on MELBO and Deep Causal Transcoding (DCT). CPE is essentially a higher-order version of the eNTK approach described above – instead of decomposing just the Jacobian, it implicitly decomposes a weighted sum of differential tensors that are higher order in weights. Taken with the eNTK work, we could ask: when are local, first-order directions enough, and when do we need finite-radius or higher-order perturbations to identify behaviorally meaningful structure? Current experiments perform CPE on only a middle slice of a transformer, although in principle could be applied to the entire transformer. CPE is meant to back-chain from alignment-relevant problems where sample-efficient feature discovery is needed. With minimal tuning and tiny data-sets (below 64 prompts), CPE can: recover 85% of password-locked performance on password-locked Llama3-70B. virtually eliminate alignment-faking on Llama3-70B-based model organisms (increasing min compliance rate across free/paid users from 2% to 92%). be used to initialize GPT-OSS-20B in an “aligned basin” in a model reward hacking environment, achieving similar oracle performance as training with oracle reward while training with the mis-specified reward . Comparative Advantages and Near-term Work. How to choose one approach over the other comes down to understanding when the local, first-order directions of the eNTK are enough and when finite-radius, higher-order heuristics like CPE are needed to identify behaviorally meaningful features. The eNTK’s eigendecomposition is the SVD of the Jacobian, which is well-understood and easier to debug than NP-hard tensor decompositions. It can also be constructed from kernels used in mean-field theory, opening a window between interpretability applications and learning theory and making the eNTK a natural candidate for forward chaining from toy model analysis. Going forward, we would like to: scale the eNTK eigenanalysis to larger models/datasets and compare their performance on feature discovery to SAEs. compare with toy models of feature learning based on mean-field theory and saddle-to-saddle dynamics, where we can carefully compare the results to higher-order methods. understand if the way that eNTK-related observables evolve throughout training could be used to build safety-relevant tools such as an early warning system for grokking or a gradient-based data attribution method. On the other hand, saddle-to-saddle and SLT theories suggest higher-order information is needed for understanding structure in neural networks. CPE searches for meaningful perturbations at a finite non-zero radius around the current network, which may ultimately be more meaningful, and does not require forward-mode autodiff kernels, making it more readily scalable to larger models. Our current work learns rank-1 LoRAs on attention outputs across several layers, softly steering towards orthogonality over the flattened adapters without the need for excessive tuning. In the future, we want to be able to include higher than rank-1 perturbations for increased expressivity, which will require a more compositional measure of diversity than simple flattening. We are also currently studying whether decomposing the loss landscape directly (as opposed to using a purely unsupervised criterion) can aid science-of-fine-tuning-type alignment research. Can we decompose independent generalizations on a real LLM? We expect this to require both a mixture of high-quality model psychology and algorithmic research. We plan to validate progress for bottom-up approaches by applying these techniques to alignment-relevant tasks. These can be broken into two broad categories: Model Auditing: Problems like backdoor/sandbagging/deception detection, data attribution, or detection of “off-target fine-tuning effects” like emergent misalignment. Model Reshaping: Studying whether perturbations in weight-space can reshape model behavior in a constructive way, particularly when standard techniques are “stuck”. For example, alignment faking is an example illustrating scenarios where standard RL may become stuck in sufficiently self-aware models due to exploration hacking, and was a successful application studied in our recent preprint . More ambitious extensions may study behavior reshaping which is more relevant to current-day practice, for instance trying to correct some of the weird generalization behaviors exhibited by gemini models , provided these problems can be reproduced faithfully enough in an open-weights setting. Top-Down Hierarchical Architectures: Scaling Sparse Transformers The bet behind this research agenda is that we can make models sparse by fiat if we guess the right ansatz and are clever about implementation. There’s nothing special about dense models from an expressivity standpoint, as evidenced by prior work on weight-sparse transformers . If anything, myriad work supporting the circuit hypothesis suggest that dense models emulate sparse, at least in an approximate, platonic sense. The only reason why dense dominates in practice is because it’s well-suited to standard hardware. Previous work estimates that it is around 100-1000x more expensive to train a weight-sparse transformer to the same level of performance as an equivalent dense model. Our recent pre-print is a first step to overcome this hurdle. The central primitive we introduce is a way to generate approximately-orthogonal hashed feature vectors. This trades memory access (slow on GPU) for compute (cheap and plentiful), which gives us the flexibility to score features and convert back to dense representations on the fly. In practice, this lets us scale to ~130k features per layer at only ~1.5x dense training throughput at the 1B parameter scale. Most of the remaining alignment tax shows up as data inefficiency (~5-10x slower than dense). The architecture only induces activation sparsity, but we demonstrate that this already yields more causally relevant directions than post-hoc SAEs. Our goal in the next 6 months is to use some of the same computational primitives (and perhaps new ones) to make the architecture more weight-sparse while staying below the current alignment tax of ~10x dense training costs. We acknowledge that this problem is hard. In order to make progress, we must be honest about the key underlying bottlenecks, and then try to reason from first principles how to address them one-by-one. We present some of these below to be expanded in an upcoming set of open problems: Overcoming the alu:mem gap . A fully sparse model must account for the time cost of moving a float from HBM to on-chip, which is around 100-1000x more than performing a FLOP with that same float. Conditionally loading only active weights of a sparse model wins compared to dense only if the active set is truly tiny. Comp-in-sup feature counting . Various results in the computation in superposition literature suggest that a ~square dense MLP taking inputs in $\mathbb R^{d_\textrm{model}}$ can perform computations on around $O(d_\textrm{model}^2)$ many sparsely-activating features operating in superposition (up to logarithmic factors in the denominator). This suggests that enforcing sparsity in standard basis for the relatively “narrow" values of $d_\textrm{model}$ used in practice (between $1024$ to $4096$) is not enough - the model will likely still operate in a superposition regime to emulate finer-grained features. Ensuring there are “Enough collisions at init”. In a very wide, sparsely activating network, feature pairs rarely co-fire, burying interaction signal below SGD's noise floor. A natural fix for this is to structure computation hierarchically, starting with a small number of more frequent, coarse-grained features that learn non-trivial interactions, and adding capacity with finer-grained features (from commonly co-occurring subsets of coarse-grained features) as needed. Maintaining Interpretability in Attention. Softmax attention can mix tokens densely, even over a sparse residual stream, particularly early in training before attention patterns have sharpened. A fully sparse transformer needs additional filtering or sparsity constraints on attention. Though challenging, this problem overlaps with standard capabilities research, with a wealth of strategies to borrow. Data Models and Validation Methods For the past several months, our focus has been exploring a data model consisting of hierarchical functions defined on critical mean-field percolation clusters embedded in a high-dimensional data space. The resulting data distribution comprises sparse, low-dimensional fractal clusters with a power-law distribution of cluster sizes. Latent variables modeling a taxonomic hierarchy generate each data point's target value. The data model is analytically tractable with known critical exponents that fix its scaling properties without requiring hyperparameter tuning. Since our last progress update , we improved the algorithm used to generate the data’s latent hierarchical structure. The previous code created undirected treelike graphs as an approximation of high-dimensional percolation clusters by growing them using a preferential attachment process. Our most recent paper , presented at the 2026 Mechanistic Interpretability Workshop at ICML 2026, replaced this with an exact procedure that leverages a mapping between percolation clusters, random trees, and additive coalescence. The code to generate synthetic datasets based on this model now implements an almost linear-time algorithm to jointly sample a random tree and its hierarchical latent decomposition, efficiently generating graphs with the precise distribution of high-dimensional percolation clusters and enabling data generation at arbitrary scale. Caption: The percolation data model. (a) Inputs are distributed as self-similar fractal clusters with power-law sizes. (b) Targets are generated by hierarchical latent variables decomposing each cluster. One important question to ask is: how does a neural network keep track of the ground truth latent hierarchy generated by a percolation dataset? Are these linearly represented within the network’s internal activations? We trained a residual MLP on our synthetic dataset and trained linear probes to regress the dataset’s latent values, grouped by depth in the latent tree. We found that the ground-truth variables are more linearly accessible in the activations of the MLP compared to the raw input, with the performance gap shrinking for larger subtrees. This outcome is consistent with the hypothesis that coarser structure is more accessible from the input geometry. The code for training neural networks on percolation data is available here . Where this is going. Building on the initial probing experiments, we plan to rigorously verify other hypotheses about neural representations using causal intervention methods. To test hypotheses about `interpretable’ features, we will also compare our latent features with those recovered by sparse autoencoders. We also intend to scale up the data generation code to produce larger datasets with more data points per latent, and to run experiments varying the width, depth, and initialization of networks trained on this data. In addition, we will release public datasets to make the percolation framework accessible to the broader mechanistic interpretability community as a tractable, synthetic sandbox for interpretability experiments. We are committed to creating better validation methods of empirical or theoretical feature hypotheses, which we think is essential for enabling ambitious interpretability of advanced AI systems. In the next year, we aim to produce a competitive benchmark to assess the performance of interpretability tools. Our goal is to use the properties of tractable but realistic synthetic datasets to design evaluation metrics with more robust faithfulness guarantees than the state of the art, enabling the reliable assessment of a tool’s ability to interpret model internals. Four questions guide future work: What structural properties of data, instantiated in synthetic datasets, replicate the behavior of deep neural networks trained on natural learning tasks? Example experiments include: Measuring transfer learning on natural data from pretraining on synthetic data. What structural properties and hyperparameters improve performance? Studying synthetic models of representational alignment by separately varying the random seeds for the latent variables and embedded graphs in the percolation model. Do natural datasets have hierarchical structure? If so, how should we measure, model, and interpret it? Example projects include: Designing an autoencoder to fit a self-similar fractal distribution and training it on natural data. Looking for evidence of hierarchical structure in data by applying hierarchical clustering methods to sparse autoencoder features. How does data structure shape the concepts used by intelligent systems? Example directions include: Defining quantitative metrics to describe how well a neural network reconstructs the latent forest in the percolation model and applying them to evaluate the features reported by interpretability tools. Understanding how the hierarchical latent variables described by the percolation model relate to existing work on concepts, including natural latents and condensation. What observable properties of datasets and trained networks differentiate data models? Examples include: Comparing the percolation kernel spectrum to the power-law spectra observed in real data. Investigating the scaling laws resulting from a data distribution consisting of a fractal cluster or a set of data manifolds with a power-law size distribution. Applying the skewed latent hierarchy described by the percolation model to investigate typicality and asymmetrical similarity in learned representations. Coming Soon We'll be sharing more about PIRAMID's work, as well as PIAMI (Physics-Informed Ambitious Mech Interp) -- a research program coordinated by PrincInt -- soon. Discuss
Score: 25🤖 ModelsJul 27, 2026https://www.lesswrong.com/posts/T2REsZneix3bmAKtL/piramid-progress-and-plans - In The Age Of AI, Trust Is A Competitive Differentiator
Customers now navigate a world where they increasingly expect to be misled, manipulated, or deceived. That raises the stakes for every AI investment organizations make.
Score: 25🌐 MovesJul 27, 2026https://www.forrester.com/blogs/in-the-age-of-ai-trust-is-a-competitive-differentiator/ - How AI is expanding what people do at work
New OpenAI research shows how AI is expanding what workers do, with ChatGPT users taking on tasks across roles and reshaping job boundaries.
- How AI-Native Service Firms Change Professional Work
AI-native service firms combine software, professionals and responsibility. Lightbringer shows why insurance, training and customer recourse matter.
Score: 25🌐 MovesJul 27, 2026https://www.forbes.com/sites/guneyyildiz/2026/07/27/how-ai-native-service-firms-change-professional-work/ - How AI Shortens Research And Innovation Cycles
AI's ability to accelerate research cycles, in and of itself, will be crucial for freeing researchers for higher-level tasks.
Score: 25🌐 MovesJul 27, 2026https://www.forbes.com/councils/forbestechcouncil/2026/07/27/how-ai-shortens-research-and-innovation-cycles/ - The AI-hooked younger generation is sourcing chatbot help even for in-person talks
Gen Z is increasingly leaning on AI chatbots to navigate dating, texting, and even face-to-face conversations, a trend experts warn could erode self-trust over time.
- LEO Technologies Launches Verus Voice AI, Giving Corrections Agencies a Defensible Way to Stop Personal Identification Number (PIN) Abuse in Real Time
LEO Technologies Launches Verus Voice AI, Giving Corrections Agencies a Defensible Way to Stop Personal Identification Number (PIN) Abuse in Real Time Toronto Star
- Is your Java runtime ready for AI attacks?
Is your Java runtime ready for AI attacks? InfoWorld
Score: 24🌐 MovesJul 27, 2026https://www.infoworld.com/article/4200319/is-your-java-runtime-ready-for-ai-attacks.html - AI cites the deep pages but sends humans to the homepage — most sites are built backward
If your business depends on people clicking through to web pages, the last two years have been brutal. Pew Research Center tracked the browsing behavior of 900 U.S. adults and found that when Google shows an AI summary, users click a traditional result just 8% of the time , roughly half the 15% rate when no summary appears. Links cited inside the AI answers themselves fare worse: users click on them only about 1% of the time. This has had a huge impact on publishers. Chartbeat data reported by Axios shows page views from Google Search fell 34% across its publisher network between December 2024 and December 2025, and small publishers have lost roughly 60% of their search referral traffic over two years. Business Insider's organic search traffic dropped 55% over three years , and some smaller publishers have already shut down. Chatbot referrals, meanwhile, still account for less than 1% of publisher page views despite growing more than 200% in a year. The story in publisher circles has been simple: AI is killing the web. But recent developments show the reality is more nuanced. Machines are reading more than ever Similarweb's 2026 Generative AI Landscape report reveals that while AI platforms send fewer humans to web pages relative to the answers they generate, the AI systems themselves are consuming the web at an accelerating rate in the form of searching the web on the user’s behalf to answer their questions. The share of ChatGPT answers containing live web citations grew more than fivefold in under a year, reaching 6.8% of all answers by May 2026. In some categories like travel, it is as high as 22.6%. Every major AI search product fetches live pages from search indexes and synthesizes answers from them, which means the quality of AI answers depends directly on the health of the content layer underneath. As Lily Ray, VP of SEO and AI search at Amsive, puts it in the Similarweb report, if your organic visibility dips, your AI search visibility follows, because the models are less likely to find your content. This has created a troublesome feedback loop. AI answers are built on an information supply chain whose funding model, ad-supported clicks, is collapsing primarily because of AI answers. A critical question for the continued viability of the open web is whether some alternative business model will work, and what the model will be. The replacement economy is forming inside the chat There is some early data showing where things may be going. Following ChatGPT's May 7 search update , which surfaced prominent clickable brand links inside answers, referral traffic from ChatGPT surged by 157% in a week. But the shape of that traffic changed: the share of referrals landing on homepages more than doubled, from roughly 25% to nearly 60%. Traditional search sent users to specific articles and deep pages tied to specific queries. AI referrals increasingly deliver a pre-informed visitor to a brand's front door. The chatbot does the researching and comparing; the human arrives ready to act. Similarweb's data shows AI-recommended brands receive two to four times as many subsequent visits as competitors that were not recommended. Money follows the behavior. Sponsored results appeared in 26% of U.S. desktop ChatGPT conversations in June 2026, up from 14% just a month earlier, per Similarweb's ad intelligence data. Two-thirds of those ads appear after the second prompt, targeted on conversation context rather than a keyword. Click-through sits around 0.50%. The traditional search engine keyword auction is being replaced by something new: paid placement inside a conversation, targeted on accumulated context. That is a direct challenge to the auction Google has run, and dominated, for two decades. Google's monopoly meets a new competitor Google is not a bystander here; it is simultaneously the incumbent being disrupted and one of the largest players in the disruption. AI Overviews now appear in a growing share of Google searches, more than 40% by May 2026 per Similarweb, and visits to Google's conversational AI Mode have climbed steadily since launch. Google is cannibalizing its own click economy rather than ceding the territory. But the ground was already shifting under Google’s core business. eMarketer projects Google's share of U.S. search advertising will fall below 50% in 2026 , the first time since roughly 2004 . The biggest chunk of that lost share is going to Amazon, whose sponsored product searches count as search advertising and which are growing three times as fast as Google's. Conversational ads are barely a rounding error in that accounting right now, but they open up a second front in a war Google has got used to not needing to fight. Meanwhile, more structural shifts are coming for Google. A federal court entered final judgment in the DOJ search antitrust case in December 2025 , imposing remedies that bar exclusive default agreements and require Google to share search data with qualified competitors. Google appealed in January 2026; the DOJ cross-appealed seeking stronger remedies. However the appeals resolve, the de facto arrangement that made Google the web's tollbooth, defaults everywhere and a closed index, is ending just as conversational advertising is changing the landscape. The competitive landscape that results is genuinely new. OpenAI, Google, Perplexity, and Microsoft are now competing not just for users but for the advertising demand that funded the open web, and none of them, including Google, controls the new surface the way Google controlled the old one. Does conversational advertising help or hurt the open web? It’s not clear whether this new model helps or hurts the web. The web as a destination for human attention is shrinking, and the ad-supported publishers built for that web are in real trouble. The web as a machine-readable substrate is growing in importance, and a new referral and advertising economy is forming that routes value to brands rather than to content pages. The problem for publishers may be that they are powerless to influence the outcome. Ahrefs, analyzing over a billion data points across its studies, found that 67% of ChatGPT's most-cited sources are things marketers cannot influence : Wikipedia alone accounts for nearly 30%. And 28.3% of ChatGPT's most-cited pages have zero Google organic visibility, suggesting the retrieval layer is only partially tethered to traditional search, a complication for anyone assuming SEO success translates cleanly. Your website needs to be rebuilt for the new way people find it According to three independent datasets, in the new world, the pages AI systems cite and the pages AI systems send humans to are different pages doing different jobs. That’s a big change, and most teams are still optimizing for the old world. Similarweb's data shows 65% of ChatGPT-cited URLs sit two or three folders deep in a site, while 58.8% of referral traffic lands on homepages. Ahrefs found the same split in its own analytics: more than 80% of its AI referral traffic goes to its homepage, product pages, and free tools , not its extensive editorial content. And a Previsible analysis of 6.77 million AI-referred sessions found a third destination: 28.8% of ChatGPT referrals land on internal site search pages, a navigation surface most publishers have long neglected precisely because Google searches were doing it for them. The right action to take is to audit your search traffic patterns. Pull your AI referral logs and whatever citation data you can access, and map which pages are being quoted as evidence versus where visitors actually enter. If it looks like you’re in the new world, there are three clear things to do: Deep pages, documentation, comparisons, and benchmarks should be structured to be citable: specific claims, clear headings, and descriptive URLs (Ahrefs found pages with natural-language URL slugs get cited at 89.78% versus 81.11% without ). The homepage should be rebuilt for a visitor who arrives with context from a conversation rather than from a blue link. They already know you have what they need, get them to it as quickly as possible. And internal search, a neglected feature on most sites, is now an acquisition surface that deserves real UX investment. There’s a lot that’s still unknown or in flux here. But the underlying shift is confirmed by every independent source that has looked: the click economy is not coming back, and the entities that learn to be quoted by machines and to convert the humans those machines send will own whatever the web becomes next.
- This startup helps brands get discovered across AI search platforms
This startup helps brands get discovered across AI search platforms YourStory.com
Score: 24🌐 MovesJul 27, 2026https://yourstory.com/2026/07/this-startup-helps-brands-get-discovered-across-ai-search-platforms - Why Enterprise Trust Is India’s Next AI Growth Frontier
Zero data retention is today a promise, not an architecture. As enterprise knowledge flows into the models, a new trust…
Score: 24🌐 MovesJul 27, 2026https://inc42.com/resources/why-enterprise-trust-is-indias-next-ai-growth-frontier/ - What is the EU’s 'Chat Control'? Ask the Euronews AI chatbot
The EU greenlit "Chat Control” to protect children online. Under this regulation, tech companies read private online messages and emails to detect child sexual abuse material (CSAM). But what is “Chat Control” and how does it work? Ask the Euronews AI chatbot.
- Y Combinator turned a sports arena into AI's biggest pep rally
Y Combinator turned a sports arena into AI's biggest pep rally Business Insider
Score: 24🌐 MovesJul 27, 2026https://www.businessinsider.com/y-combinator-startup-school-arena-2026-7 - HSBC to hire 100 AI specialists, 100 wealth managers in boost to Singapore hub
HSBC to hire 100 AI specialists, 100 wealth managers in boost to Singapore hub Reuters
- FOD#160: What Is Latent Reasoning? How AI Can Think Without Words
Latent reasoning moves AI computation from token-based CoT into continuous hidden states. It could make reasoning faster and more flexible, while removing one of the few parts humans can still inspect
- AI Found a New Way to Ruin Memes: Selling Them to Advertisers
The creator of the 'Running Away Balloon' meme is suing a company for offering the copyrighted comic as a template for advertisements.
Score: 22🌐 MovesJul 27, 2026https://gizmodo.com/ai-found-a-new-way-to-ruin-memes-selling-them-to-advertisers-2000791358 - Cory Doctorow on using AI without letting AI use you
Award-winning author and tech activist is back with new book, “The Reverse Centaur’s Guide to Life after AI.” The post Cory Doctorow on using AI without letting AI use you first appeared on BetaKit .
- The evolving role of network engineers in the age of AI
AI is reshaping networking: integrated security, real-time visibility and predictive automation help engineers prevent issues before disruption.
Score: 22🌐 MovesJul 27, 2026https://www.techradar.com/pro/the-evolving-role-of-network-engineers-in-the-age-of-ai - Designing Human–AI Systems
Designing Human–AI Systems Cambridge Advance Online
- Reducing Human Annotation with ML Active Learning
In a world where human time is expensive, learn how to use it only when really necessary The post Reducing Human Annotation with ML Active Learning appeared first on Towards Data Science .
Score: 22🌐 MovesJul 27, 2026https://towardsdatascience.com/reducing-human-annotation-with-ml-active-learning/ - Why a 'Clean' Property File Isn't Enough: How Cubictree's AI Protects Lenders from Hidden Fraud
Why a 'Clean' Property File Isn't Enough: How Cubictree's AI Protects Lenders from Hidden Fraud YourStory.com
- What If We Got AI Right? by Eleanor Drage review – avoiding apocalypse
The academic’s ambitious guide to the ethics of tech falls short of its promise to provide meaningful answers The AI ethicist Eleanor Drage believes that to thrive alongside artificial intelligence, humans need to recognise AI’s “humanity”. She means that while we often speak of AI as though it were some mystical, formless thing, it is the product of hours of human labour, in silicon and quartz mines, microchip factories and data labelling centres. If citizens want to wrest power from the tech companies that are building AI, it helps to understand clearly what they are selling us. This also means seeing through the marketing and hype: “the cloud”, that nebulous, distant place where your photos are stored, is just someone else’s computer; an AI “hallucination” is a system error or data labelling mistake; intelligence is a very broad – and potentially misleading – term. The story in Silicon Valley is that AI will either save us, ushering in a new era of prosperity and human longevity, or it will wipe us all out. Drage argues, convincingly, that big tech cannot possibly deliver on its utopian promises while it continues to chase profit and power above all else, and while it serves the interests of a narrow few. Meanwhile, the cultural obsession with the AI-induced apocalypse conveniently redirects attention from the practical conversations we could be having about how to make AI safer – by ensuring citizens have more power over how their information is used, and how AI models are trained and regulated, for example. Tech titans’ apocalyptic visions also create the false impression that AI safety is more important than climate change, geopolitical tensions and inequality, three problems that are already causing a huge amount of suffering and that, if addressed, would reduce many AI-linked risks too. Continue reading...
- Why The Biggest Opportunity In AI Isn’t Automation, It’s Visibility
Technology creates consistency. People create connection. The strongest customer experiences require both.
- Waymo Is Racking Up Thousands of Dollars of Parking Fines in Austin
As robotaxis scale, cities and law-enforcement agencies are cracking down on bad behavior.
- Research focuses on Rwanda to help root out online propaganda in languages overlooked by AI
Research focuses on Rwanda to help root out online propaganda in languages overlooked by AI EurekAlert!
- AI underwater robots can now track diver stress via exhaled bubbles
University of Minnesota Twin Cities researchers have developed a first-of-its-kind AI system that allows underwater companion robots to monitor a diver's health in real time simply by "watching" their exhaled bubbles.
Score: 21🌐 MovesJul 27, 2026https://techxplore.com/news/2026-07-ai-underwater-robots-track-diver.html - From Aisle to AI: Industry leaders tackle GLP-1, retail media and AI-powered shopping trips at Groceryshop 2026
From Aisle to AI: Industry leaders tackle GLP-1, retail media and AI-powered shopping trips at Groceryshop 2026 Toronto Star