AI News Archive: July 20, 2026 — Part 8
Sourced from 500+ daily AI sources, scored by relevance.
- IssueBench - How We Evaluate Engine
A guide on evaluating engine performance with IssueBench.
- Oxford Leading AI Adoption in Financial Services Programme
Oxford Leading AI Adoption in Financial Services Programme University of Oxford Saïd Business School
Score: 28🌐 MovesJul 20, 2026https://www.sbs.ox.ac.uk/oxford-leading-ai-adoption-financial-services-programme - DaVita puts connected data to work for kidney care
DaVita puts connected data to work for kidney care Healthcare IT News
Score: 28🌐 MovesJul 20, 2026https://www.healthcareitnews.com/news/davita-puts-connected-data-work-kidney-care - This $80 AI dictation app could replace hours of typing
Save time with Voibe, an offline AI voice dictation app for Apple Silicon Macs that transcribes speech up to three times faster than typing. Lifetime access is $79.99 for a limited time.
- I used Google AI Mode to create new YouTube Music playlists instantly — and these 7 prompts helped me find new music faster than ever
I used Google AI Mode to create new YouTube Music playlists instantly — and these 7 prompts helped me find new music faster than ever Tom's Guide
- UK data center startup Nscale bets big on Bellevue for U.S. engineering hub amid AI boom
AI infrastructure company Nscale is expanding its presence in Bellevue with a new U.S. engineering headquarters, adding to the Eastside's growing role as a hub for artificial intelligence talent and cloud computing infrastructure. Read More
- I recreated OpenAI's Codex Micro for a lot less - and mine is more customizable
The $230 Codex Micro sold out before I could buy one, so I turned my Stream Deck+ into a surprisingly capable alternative.
Score: 26🌐 MovesJul 20, 2026https://www.zdnet.com/article/codex-micro-keypad-alternative-stream-deck-plus/ - Chiemela Uzunma Kalu, Recognized by Influential Women, Advances AI-Driven Innovation in Higher Education
Chiemela Uzunma Kalu, Recognized by Influential Women, Advances AI-Driven Innovation in Higher Education azcentral.com and The Arizona Republic
- Rethinking local structural awareness in graph representation learning - ORA
Rethinking local structural awareness in graph representation learning ORA - Oxford University Research Archive
- Tempus AI Coverage
Tempus AI Coverage MedCity News
- CGS Immersive Introduces Cicero Analytics, the Observability Layer Connecting Workforce Readiness to Business Performance
CGS Immersive Introduces Cicero Analytics, the Observability Layer Connecting Workforce Readiness to Business Performance Toronto Star
- The psychology of believing AI: Certainty vs truth
A client sits down, starts to speak, and within a minute, I can hear that the problem has already been shaped somewhere else. The language arrives too clean; the situation has an outline before the two of us have looked at it together. Later in the session, they often tell me where the outline came […] The post The psychology of believing AI: Certainty vs truth appeared first on e27 .
Score: 25🌐 MovesJul 20, 2026https://e27.co/the-psychology-of-believing-ai-certainty-vs-truth-20260719/ - Q&A: Why boutique consultancies might be better for AI rollouts than the bigwigs
Major AI labs are unleashing forward-deployed engineers (FDEs) to try and grab enterprise customers. Large consultancies are dishing out tokens and assembling armies of consultants — both human and agent — to do the same. But smaller firms are in the mix now, as well. AI is helping 28Stone Consulting , a New York-based, 230-person technology consultancy for capital markets, punch above its weight against larger rivals in the rush to deliver FDEs . In this Q&A, Thomas Dolan and Frank Erickson , founders of 28Stone, argue that agentic AI isn’t a one-size-fits-all solution in vertical markets; success takes discipline, deep domain expertise, and human involvement to mitigate risk. Many enterprises continue to struggle with the use of AI agents, which is consultancies are stepping in to get projects off the ground. 28Stone is among those that have published blueprints and methodologies on the development and delivery of agentic AI workflows with humans in the loop. Computerworld spoke with both founding partners about why companies are still stumbling with agentic AI rollouts , and what a disciplined delivery process actually looks like. After 15 years of delivering software for capital markets firms, is ‘AI-first’ a real distinction or just positioning? Dolan: “We’re not shying away from being AI-forward. What needs to shine through is AI done intelligently — not stuff you get by buying some tokens for somebody on the trading desk. We’re an AI-first firm.” Erickson: “And it’s temporary. At some point, AI is going to be synonymous with software development. “The whole idea of an AI SDLC (software development lifecycle) versus an SDLC is going to be one and the same, a lot like cloud computing today. To not include AI in your strategy, you’d look like a COBOL vendor.” What does agentic AI delivery look like? Dolan: “We’ve got several AI initiatives delivering a pure agentic approach. We’ve doubled down on the human expertise wrapper in the SDLC. That doesn’t mean sacrificing any of the benefits of the AI models — quite the opposite. “You don’t achieve anywhere near the same level of value from applying AI without keeping that expertise — industry, functional and technical — throughout the process.” Where do humans stay in the loop once agents are doing the work? Dolan: “We’re believers in starting with requirements discovery. Someone who knows the analytical nuances of a good business analyst is critically important; shaping a product owner’s business information through a markup file that can be fed into a BA agent, then treating the output as if it came from a very fast junior BA. Only then is the story complete. “The developer takes that story, transforms it into the most efficient input, then owns the output, because they’re accountable for that code. A developer should own the code on both the input and output side. “Your product owner, who knows the business, that’s great. But expecting them to interact with an agent and output enterprise code is ridiculous. It’s not a great plan.“ Why not just put one do-everything person in charge of AI and agents? Erickson: “Every analyst, programmer or software engineer isn’t a great requirements analyst. And a great domain analyst with some technical background won’t know if the agent’s code is garbage, maintainable, performant. “It’s unrealistic to expect one individual to have that breadth across domain, software engineering, testing, deployment. Clients ask all the time, and we push back: ‘Great, if you can find that guy, they’re few and far between.’ To deliver at the enterprise level, you need the human expertise, at depth.“ Dolan: “There’s system speed and latency, important in parts of finance. Then there’s speed of delivery, because other areas evolve quickly and time-to-market is critical. “Our human wrapper may at first pass come across as a little slowed down. Maybe it is. But [Erickson] has a good analogy about one of the dangers of AI: you can end up going really fast in the wrong direction. By the time you look up, you’re way off base and have to backtrack.“ What about AI in your sector do you think is overhyped? Dolan: “The hype around the ease of use of AI and the democratization of enterprise software delivery — that ‘anybody could do it now, it’s all being done by machines’ — is another idea that could prove costly in the long run. “This do-it-yourself reaction is dangerous for clients, and for trust in the overall AI benefit, which is real. We compare it to the beginning of offshoring 20, 30 years ago: a golden idea that was going to cure everything. A lot of firms did it thoughtlessly, thinking it’s just labor arbitrage, and it almost inevitably failed. That all-or-nothing mentality missed that offshoring is an amazing way of getting better value for your dollar, but it has to be done thoughtfully, so the delivery process — the thing that ties it all together — stays unsevered. “We’re seeing that now. I’ve heard, ‘We’ll just push a button, the machine’s building the system.’ The machine is not building the system. It might be writing the code, the story, running the tests. The system is built by a team of engineers you bring in and trust. My fear is that people will say, ‘We don’t need this vendor or this technology team. I’ve got a product team. They might not be able to code at all, but they know the business,’ and it fails dramatically. “Then people say, ‘We played with AI, it’s not ready yet,’ and throw it all away. One of the best things we can do is ensure clients know the benefit is real.“ Erickson: “The hype can be summed up in a single phrase: vibe coding . That has done AI a massive disservice, because there’s a huge difference between vibe coding and enterprise software development, and some of the loudest proponents of AI are too latched on to it. In our industry, the only way to succeed would be a stable of unicorns. It just doesn’t scale. I get perturbed when our people internally refer to AI tooling as vibe coding; if they think that’s what they’re doing, they’re misunderstood.“ When you engage clients at different levels of AI maturity, how do you get them to a understand what works? Dolan: “95% of our take on an agentic approach is in line with everyone else’s, but that 5% matters, especially in requirements discovery, in who’s giving the requirements and how they’re thought of. It can set you up for dramatic errors, given the speed at which you’re moving. “There’s a dangerous human tendency we’re seeing among clients to try and cut corners at the start of a project and — in lieu of having deep, expert driven discovery sessions — just summarize what they may want using AI. “We would hope our clients are collaborative, everyone understanding it’s early days. If a client insists on doing something we feel strongly against, like a product owner completely owning everything right up to code generation, that’s an issue we have to either push back strongly on or step out of the accountability for.“ AI body shops — LLM providers and giant consultancies — are emerging to help enterprises deploy AI. Does that model work? Dolan: “Whether you’re partnering with an LLM or with an AI-first, generic software provider — ‘Hey, we’re not industry guys, but we know AI delivery’ — you end up, if you’re a bank or a broker-dealer, saying: ‘All right, we know our business, these guys know the AI side of it. What could go wrong? Put us together and we’ll have quality engineering.’ “The problem is what you miss: the know-how of putting industry and technical expertise together and actually delivering financial services systems. The people working at the generic delivery firms, whether an AI-only firm or a body shop somewhere, don’t have that capability.“ Does AI change the economics for smaller consultancies like yours competing against the big firms, and does it cut both ways? Dolan: “Over our 15 years pre-AI, there were two recurring reasons we’d lose a project. One: ‘We’d love to work with you guys, given your subject matter expertise, but the costs just aren’t there compared to my budgets. I’m being forced to go to a body shop or an [offshore] delivery center.’ The other side of that coin: ‘We love your capabilities, but you’re a firm of 230 people and I need 300, 400 people.’ “AI changes the options for clients. You don’t have to sacrifice the niche vendor who knows your space just because you need a larger team or a cost target. AI levels the playing field and should allow smaller firms to compete with the larger, big-box generic firms, the Accentures of the world.“ Erickson: “It redefines what scale means. You can look at velocity as a measure of your cost to deliver, not a rate card. Scale can’t be defined in terms of headcount anymore. It’s got to be defined in terms of output. “There’s a threat in it, too. If you’re an Accenture with hundreds of thousands of low-cost software engineers, how do you train all those people? I feel for them. But for us, a couple hundred people with a specific domain focus, it’s a huge opportunity.“ How has the profile of the people you and others hire changed with this agentic process? Erickson: “You’re still looking for people with strong engineering and design backgrounds, and communication skills, because they interact across the software development lifecycle more than in the past. “Many take too much joy in typing out perfect code. Sorry, I don’t need you writing for-loops and classes anymore. I need you reviewing them, understanding them, operating at a higher level. That’s a different kind of person: an engineer, not a programmer or a coder. On the [business analyst] side it’s similar: people took great pride in detailed user stories covering every path. Now it’s conversations, prompts, reviewing output — less doing, more interacting. “More than ever, they have to be interested in the domain. They can’t just be, ‘I want to learn everything there is to know about Java.’ That’s too narrow. They don’t have to be an expert; they have to be interested. In our case, capital markets is a specific niche. The biggest challenge is getting familiar with the tools — finding time, while delivering for customers, to ramp up and make the mistakes you need to without jeopardizing projects.“ What about governance? Who’s keeping AI delivery and its costs under control? Erickson: “This is evolving rapidly. People aren’t sure how to put governance around this. The most obvious is financial governance. People are starting to get hefty bills. One of our clients spent a million dollars on tokens over the last eight weeks alone. Sticker shock. The token-maxing policies are starting to show their flaws. It’s wild west still: learn on the fly, then figure out what needs to be governed.“ Are CIOs actually opening their wallets? And when they do, what’s the smarter way to invest? Erickson: “There’s still a lot of caution. Forecasts keep going down on how long something should take. So: ‘I could wait three months and maybe still get it delivered by the same date someone’s promising me now, but for half the price. I’m going to wait and see when equilibrium is met.’ We haven’t seen the wallets open up like crazy — it’s slow adoption.“ Dolan: “One of our clients is looking at it from a productivity-boost perspective: instead of doing the same for less, I can do much more for the same. AI lets clients pull the trigger on things they wouldn’t have in the past — projects that might not have been approved pre-AI, where the costs have come down to a point that’s palatable with the business.“ Erickson: “And that’s the story we’re hoping to hear more of. There isn’t a huge cost anymore to exploring a business opportunity. The time and money that would have gone to a return-on-investment study could be spent on a proof-of-concept with AI, and the project done a few weeks later. Maybe [there’s] a hint of things to come, where decisions start being made quicker. “There’s a little fear on our side, though: a lot of tiny little projects is tough for a consulting business.“
- Your iPhone Messages app has a real-time translation feature — here's how to use it
Your iPhone Messages app has a real-time translation feature — here's how to use it Tom's Guide
- I found the 7 easiest Claude connectors for people over 50 — start with these
I found the 7 easiest Claude connectors for people over 50 — start with these Tom's Guide
Score: 23🌐 MovesJul 20, 2026https://www.tomsguide.com/ai/claude/i-found-the-7-easiest-claude-connectors-for-people-over-50-start-with-these - Songive Launches AI-Powered Platform That Turns Personal Stories Into Custom Songs in Minutes
Songive Launches AI-Powered Platform That Turns Personal Stories Into Custom Songs in Minutes USA Today
- The Download: AI hiring biases, and weather data sabotage
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. AI is more likely than humans to form biases when hiring The next time you apply for a job, AI may screen your résumé before any human sees it. But there’s…
Score: 23🌐 MovesJul 20, 2026https://www.technologyreview.com/2026/07/20/1140664/the-download-ai-hiring-biases-weather-data-sabotage/ - From instinct to real-time insight: Transforming steel sales with AI
ArcelorMittal Brazil partnered with McKinsey to redesign its sales journey. A gen AI–powered tool removed the friction of everyday selling, freeing salespeople to focus on what mattered most: creating value for customers.
- How do the new rules boost air passenger rights? Ask the Euronews AI chatbot
Air passengers will enjoy stronger protection starting mid-2027. On 7 July, MEPs endorsed the new rules with a vote of 646 to 12, maintaining current rights and adding new protections. The Council approved the measures on 13 July. But what do these include? Ask the Euronews AI chatbot.
- DropPR.ai Says Structured PR Distribution Is a Missing Layer in Most Brands’ AEO Strategy
DropPR.ai Says Structured PR Distribution Is a Missing Layer in Most Brands’ AEO Strategy USA Today
- RDB Consulting pioneers AI-driven infrastructure optimisation tools for SA Stud Book
The company has assisted SA Stud Book in improving the speed and reliability of its mission-critical infrastructure and establishing the foundation for in-house digital development.
- Madurai turns to AI and GPS to improve garbage collection
Madurai turns to AI and GPS to improve garbage collection YourStory.com
- Know How AI Platforms See Your Brand with BeeSeen for Life for $99.99
Improve your brand’s visibility and overall GEO with a lifetime subscription to BeeSeen for almost $400 less. The post Know How AI Platforms See Your Brand with BeeSeen for Life for $99.99 appeared first on TechRepublic .
- Why .AI domain names have become the digital identity of the AI era
As artificial intelligence becomes embedded across virtually every industry, digital branding decisions are becoming increasingly strategic.
- Nick Sirianni turned to ChatGPT to keep his message to Eagles fresh
Nick Sirianni turned to ChatGPT to keep his message to Eagles fresh Eagles Wire
- CanadianSME Small Business Summit 2026 Returns to Toronto with “Beyond AI” Focus on Building Intelligent, Resilient & Human-Centered SMEs
CanadianSME Small Business Summit 2026 Returns to Toronto with “Beyond AI” Focus on Building Intelligent, Resilient & Human-Centered SMEs Toronto Star
- Factor Graphs FTW: From SLAM to Convex Relaxations and Constraint Manifolds
A:Frank Dellaert (Georgia Institute of Technology); TT:Guest lecture; RL:Language, Speech & Vision;
Score: 20🌐 MovesJul 20, 2026https://ai.kuleuven.be/events/factor-graphs-ftw-from-slam-to-convex-relaxations-and-constraint-manifolds - How to check if ChatGPT and other AI tools cite your website - and improve your chances in 2026
SEO is not dead. Here's how to get cited more often and boost search visibility in AI tools today.
Score: 20🌐 MovesJul 20, 2026https://www.zdnet.com/article/how-to-check-if-chatgpt-ai-tools-cite-website-improve-chances/ - Five AI Terms Your Board Will Ask About This Quarter
5 AI trends (and terms) that affect business. Digital Labor, Agentic Lifecycle, AI Tokenomics, Forward Deployed Engineering, and Taste. What you should know.
Score: 20🌐 MovesJul 20, 2026https://www.forbes.com/sites/nishatalagala/2026/07/20/five-ai-terms-your-board-will-ask-about-this-quarter/ - Dear Microsoft: Stop sticking your AI in my IDE
Dear Microsoft: Stop sticking your AI in my IDE InfoWorld
Score: 19🌐 MovesJul 20, 2026https://www.infoworld.com/article/4197387/dear-microsoft-stop-sticking-your-ai-in-my-ide.html - Fugitive tries to wipe criminal record using AI-taught skills
Fugitive tries to wipe criminal record using AI-taught skills The Telegraph
Score: 19🌐 MovesJul 20, 2026https://www.telegraph.co.uk/world-news/2026/07/20/thai-fugitive-arrested-chatbots-erase-police-record/ - Acer Swift 16 AI (2026) Review: A Great Big-Screen Laptop With One Giant Hurdle
Acer Swift 16 AI (2026) Review: A Great Big-Screen Laptop With One Giant Hurdle PCMag
- Cabinet Boost Enhances Kitchen Cabinet Marketing Platform with Advanced AI Automation
Cabinet Boost Enhances Kitchen Cabinet Marketing Platform with Advanced AI Automation USA Today
- IT Essentials: Hype derails the AI Express
IT Essentials: Hype derails the AI Express Computing UK
Score: 18🌐 MovesJul 20, 2026https://www.computing.co.uk/column/2026/it-essentials-hype-derails-the-ai-express - Executive Interview: Vigilant AI.ai
Greg Coleshill, CCO at Vigilant AI.ai, tells CB Insights how they view the market, customer needs, and their company. How do you define your market and where does your company fit into that space? Vigilant AI sits at the intersection … The post Executive Interview: Vigilant AI.ai appeared first on CB Insights Research .
- CEO Interview: Unsiloed AI
Aman Mishra, CEO of Unsiloed AI, tells CB Insights how they view the market, customer needs, and their company. How do you define your market and where does your company fit into that space? We are building vision models to … The post CEO Interview: Unsiloed AI appeared first on CB Insights Research .
- 5 ways to build a side hustle with Gemini
An illustration of a person sitting in a chair uploading files, and an AI sparkle icon
Score: 17🌐 MovesJul 20, 2026https://blog.google/products-and-platforms/products/gemini/launch-business-with-gemini/ - With AI, activity is not value
The emergence of artificial intelligence is beginning to expose a profound weakness in the way modern enterprises measure performance. For decades , business evaluation systems have been built around the logic of the industrial and transactional economy. Revenue growth, operating margins, earnings per share, labor productivity, return on investment and market share became the dominant indicators of organizational success because they reflected the economic realities of a world in which value creation was primarily tied to physical production, labor efficiency, scale and later the automation of information processing. AI, however, is altering the very structure of enterprise value creation, and in doing so it is creating a widening separation between perceived future value and actual realized economic performance. Much of the current discussion surrounding AI performance measurement reflects this tension. The overwhelming majority of AI-related metrics being celebrated today are not direct measures of realized enterprise outcomes. They are largely indicators of capability formation, market positioning, experimentation or investor signaling. Metrics such as AI spending levels, number of AI use cases, GPUs deployed, copilots implemented, models placed into production, AI hiring growth or agentic AI pilots all serve primarily as proxies for anticipated future advantage. These indicators may influence stock valuations, analyst sentiment and strategic narratives, but their relationship to measurable operational performance is often indirect, delayed or in some cases entirely speculative. This distinction is critically important because capital markets have historically rewarded the expectation of technological transformation long before actual economic results materialized. During previous technological revolutions—including electrification, enterprise resource planning, the internet, cloud computing and mobile platforms—valuation expansion frequently preceded measurable productivity gains by many years. The market priced future possibility before operational economics caught up. In many instances, investors rewarded firms simply for appearing strategically aligned with the dominant technological shift of the era. AI appears to be following a similar trajectory. The phenomenon resembles the famous productivity paradox articulated by economist Robert Solow, who observed that “you can see the computer age everywhere but in the productivity statistics.” AI today is visible everywhere: in investor presentations, earnings calls, technology conferences, product announcements and boardroom strategies. Yet in many industries, its measurable contribution to enterprise productivity, profitability or economic resilience remains difficult to isolate with precision. This does not necessarily mean AI lacks value. Rather, it reflects the reality that traditional accounting and performance systems were never designed to measure the forms of value AI increasingly produces. Artificial intelligence creates benefits that are often diffuse, cumulative and difficult to attribute directly to financial outcomes. AI may improve forecasting accuracy, reduce fraud, accelerate decision cycles, augment employee effectiveness, improve customer interactions, optimize logistics or enhance cybersecurity resilience. These benefits frequently manifest as second-order effects distributed across the enterprise rather than as immediately visible financial events. The causal chain between AI investment and realized business performance can therefore become extraordinarily difficult to quantify. A company may become operationally more intelligent without immediately becoming measurably more profitable. At the same time, AI introduces a profound danger: organizations may increasingly optimize for technological narrative rather than durable enterprise economics. Many firms today are pursuing AI primarily because markets reward the appearance of AI leadership. Investor enthusiasm, analyst pressure and competitive fear create incentives to demonstrate visible AI activity regardless of whether measurable economic value has actually been achieved . In this environment, AI metrics can easily become instruments of valuation signaling rather than instruments of operational truth. This distinction between signaling and substance may become one of the defining economic challenges of the AI era. An organization may announce aggressive AI deployment programs, reduce headcount and report short-term margin improvements while simultaneously increasing hidden forms of technological fragility. Infrastructure costs may rise dramatically as GPU consumption, cloud usage, data engineering requirements and cybersecurity complexity expand. Technical debt may accelerate as AI-generated code proliferates without sufficient architectural discipline. Institutional knowledge may erode as organizations become excessively dependent on opaque models and automated systems. Long-term innovation capacity may weaken if enterprises divert disproportionate resources toward maintaining internally generated AI systems rather than building new strategic capabilities. What measuring AI value might actually look like The distinction between AI activity and AI value becomes clearer when viewed through the kinds of measures organizations choose to track. Many enterprises today emphasize indicators such as the number of AI models deployed, copilots implemented, agents created, prompts executed, tokens consumed or employees using AI tools. These metrics demonstrate adoption and technological activity, but they reveal relatively little about whether AI is producing meaningful business outcomes. Measures of enterprise value look quite different. A manufacturer might evaluate whether AI improves demand forecasting accuracy enough to reduce inventory carrying costs or stockouts. A financial institution might measure whether AI meaningfully lowers fraud losses, accelerates loan processing or improves regulatory compliance. A healthcare provider could assess reductions in administrative burden, faster clinical decision support or improvements in patient throughput. In each case, the objective is not simply to measure AI deployment, but to determine whether AI creates measurable improvements in operational performance, economic outcomes or organizational resilience. Ultimately, organizations may need to ask a different question: not “How much AI are we using?” but “How much business value does each unit of AI investment create?” That shift—from measuring technological activity to measuring economic outcomes—may become one of the defining management disciplines of the AI era. Under traditional accounting frameworks, many of these deteriorations remain largely invisible. Quarterly earnings may improve even as underlying enterprise resilience declines. Stock prices may rise even as operational complexity becomes increasingly unsustainable. In this sense, the AI era threatens to widen the gap between financial appearance and organizational reality. This is why the future of enterprise measurement cannot simply involve adding AI metrics to existing financial scorecards. The challenge is far deeper. AI forces a reconsideration of what business performance actually means. Historically, enterprises were measured largely through static indicators of efficiency and output. Increasingly, however, competitive advantage may depend less on traditional efficiency and more on adaptive intelligence: the ability of an organization to learn faster, make better decisions, integrate human and machine capabilities effectively, manage technological complexity sustainably and convert computational power into durable economic outcomes. The most important future performance measures may therefore revolve around questions traditional accounting rarely addresses. How effectively does an enterprise convert technology investment into sustainable business capability? How economically efficient are its AI operations relative to the value they generate? How resilient is the organization to AI failure, cybersecurity disruption or infrastructure inflation? How successfully does it preserve and amplify human expertise rather than simply eliminate labor? How rapidly can it learn, adapt and operationalize new knowledge? These are not merely technology questions. They are questions of enterprise economics, organizational sustainability and long-term competitive viability. The companies that ultimately succeed in the AI era may not be those with the largest AI budgets, the greatest number of pilots or the most aggressive automation programs. They may instead be the firms that best understand the economics of technological capability itself: organizations capable of balancing innovation with resilience, automation with human augmentation and technological ambition with sustainable operational design. The coming decade is therefore likely to produce a widening divide between enterprises optimizing for AI-driven valuation narratives and enterprises optimizing for measurable, durable economic performance. In the short term, these may appear to be the same thing. Over time, however, the distinction will become increasingly visible. Some organizations will discover that AI has enhanced genuine enterprise capability. Others will discover that they merely optimized the appearance of transformation while silently accumulating new forms of economic and operational risk. Artificial intelligence is not simply changing business operations. It is exposing the inadequacy of many of the measures used to evaluate business success itself. The central challenge of the AI economy may ultimately become not whether organizations adopt AI, but whether they can distinguish between technological activity and actual economic value creation. This article is published as part of the Foundry Expert Contributor Network. Want to join?
- From running a successful social media platform to showing a proprietary robot dog to the PM: the ShareChat founders' second act
From running a successful social media platform to showing a proprietary robot dog to the PM: the ShareChat founders' second act YourStory.com
- The Marbella Directory Launches AI Visibility Platform for Local Businesses
The Marbella Directory Launches AI Visibility Platform for Local Businesses USA Today
- Personal training app Flex AI is shutting down
CEO cites difficulty creating a sustainable business in the crowded fitness tech category. The post Personal training app Flex AI is shutting down first appeared on BetaKit .
- AI’s problems aren’t what you think
The biggest and loudest prediction about AI is that it will eliminate millions of jobs. It is dramatic and easy to repeat. But from what I’ve seen, inside most enterprises the more immediate problem has turned out to be something else entirely: a growing mass of tools, agents, models and usage costs spreading faster than most organizations can govern or connect to real business value, also known as AI sprawl. None of that invalidates the initial fear. Indeed, AI can clear backlogs, speed up analysis, draft usable content and reduce time spent on repetitive work. In my opinion, what goes wrong is the assumption that those gains will scale seamlessly, and that more AI will automatically produce more value. What really matters is not only how much AI a company can deploy, but whether its use fits inside a growth strategy, an operating model and an organization that can use it well. The early results of AI use made the logical progression feel obvious, even a foregone conclusion. If it could already improve output in narrow use cases, then broader deployment should produce broader gains. Simple! Better models were expected to deliver better results. More agents were expected to drive more automation. For many companies, this logic held for long enough to encourage overexpansion. But this logic has started to break down as usage continues to scale. I’ve seen returns diminish much quicker than expected. To illustrate, one industry analysis found that developers who used AI most heavily produced about twice the output of moderate users, but consumed roughly ten times the compute. At a certain point, more AI does not create proportionally more value – it simply becomes more expensive. But where, exactly? From experimentation to sprawl Experimentation played a key role in this downturn, but it’s not the culprit. As AI continues to sprawl, the problem continues that AI is spreading faster than most companies can coordinate. Teams often solve the same problem in parallel, paying for overlapping capabilities and layering new tools atop existing ones without any clear inventory of what ‘s already in place. What can appear as momentum is really turning into redundancy. I’ve seen versions of this play out repeatedly. At one financial firm, several business units were pursuing AI projects aimed at automating research and reporting. Each team moved independently; selecting their own tools, building their own workflows and creating separate data pipelines, with little to no coordination between teams. In some cases, different groups were developing nearly identical capabilities without realizing it, solving the same problems twice without any shared visibility into each other’s work. Individually, the projects showed real promise. Collectively, the projects created duplication, fragmented data and inconsistent standards business and enterprise wide. By the time leadership stepped back to assess, the company found itself paying for overlapping capabilities, maintaining multiple versions of the same underlying data, and struggling to determine which solutions were actually delivering value versus which were simply consuming budget and eating at engineering time. Perhaps most troubling: nobody at the enterprise level had a complete view of what was being built, by whom or why. What began as healthy, well-intentioned experimentation had, without anyone deciding it should, evolved into full-blown AI sprawl, creating a patchwork of disconnected initiatives that was difficult to govern, harder to secure and far more expensive than a coordinated approach could and should be. Early wins encourage a still wider rollout, but many organizations expand usage before they put real controls in place. Experimentation becomes sprawl. Budgets grow quickly, and few leaders have a reliable view of who is using what or why. AI strategy cannot sit beside growth strategy This is where I see many companies still get the issue wrong. They treat AI and growth strategy as two separate efforts, then wonder how adoption gets so messy. A business cannot drop AI into its operations and expect momentum to take over. The technology has to support a clear path to growth, whether that means improving margin, speed, service, capacity or decision-making. At the same time, growth plans cannot assume AI changes nothing about delivery, design or operating leverage. The real challenge is in ensuring the two work together. Personally, I’ve seen better results when AI initiatives are tied to a specific business objective from the beginning, rather than launched as broad, abstract or transformative effort. One mattress retailer I’ve worked with took this approach, starting with a single, focused and well-defined use case rather than trying to transform or overhaul the entire organization at once. The company introduced an AI-powered training platform for store associates, giving employees a low-pressure way to practice sales conversations and product recommendations before applying them to external situations with customers on the floor. Because employees experienced immediate and tangible value from the tool, adoption spread quickly across locations, with minimal need for top-down mandates. Early, visible success helped to build internal credibility and generate momentum, which leadership then leveraged to expand into more complex AI initiatives across areas such as inventory management, demand forecasting and replenishment planning. Ultimately, the technology succeeded not because it was innovative for its own sake, but because it was connected to a larger growth strategy: improving sales effectiveness on the floor, driving operational efficiency behind the scenes and strengthening workforce capability at entry level. A major lesson we walked away with here was that starting small and specific, with a clear throughline to business value creates a strong foundation for sustainable and scalable AI use. What implementation actually takes All this takes more than a few easy guardrails. It takes strategy. Leaders need a real inventory of the tools, agents and assistants already in use across the business, who owns them, what data they can access and everything that they support. They also need financial controls that match the economics of token-based usage, including role-based access, thresholds and review processes that make spend visible before it becomes a surprise. Similarly, they need metrics that go beyond mere activity. More prompts do not mean more value. If a deployment cannot be tied to throughput, margin, quality, cycle time or another tangible result, it is still unfinished. This is also why blunt shutdowns rarely work. If leaders clamp down too hard, employees often move to unsanctioned tools and create a larger shadow AI problem, or the unauthorized use of artificial intelligence tools, models or chatbots by employees, without the knowledge or approval of IT and security teams, with even less visibility and more risk. The better answer is disciplined adoption: clear ownership, rules, metrics and enough flexibility for teams to use AI where it works. That matters for the people as much as it does for the budget. Those that modernize well end up with better work – not just less of it. The story of the moment isn’t about AI replacing people – or even AI in general. It’s about whether companies know their own businesses well enough to keep incorporating powerful new tools without mistaking activity for progress. As technological capabilities continue to appear, the winners will be the organizations that understand where it belongs, what it can improve and how to turn each new wave into something permanent. This article is published as part of the Foundry Expert Contributor Network. Want to join?
Score: 15🌐 MovesJul 20, 2026https://www.cio.com/article/4198475/ais-problems-arent-what-you-think.html - How two Cambridge data science learners built an AI startup
How two Cambridge data science learners built an AI startup pace.cam.ac.uk
Score: 15🌐 MovesJul 20, 2026https://www.pace.cam.ac.uk/news/how-two-cambridge-data-science-learners-built-ai-startup - Brides and grooms are using AI to write speeches. Can you spot the difference?
Brides and grooms are using AI to write speeches. Can you spot the difference? The Telegraph
Score: 15🌐 MovesJul 20, 2026https://www.telegraph.co.uk/family/relationships/bride-groom-using-ai-wedding-speech/ - Cosmoserve Says Soft Robotic Capture Demo Gathered Key Data Despite Glitch
Spacetech startup Cosmoserve reported a partial success for its maiden Mission Embrace, saying its soft robotic capture system generated valuable…
Score: 14🌐 MovesJul 20, 2026https://inc42.com/buzz/cosmoserve-says-soft-robotic-capture-demo-gathered-key-data-despite-glitch/ - GAPVelocity AI launches VELO for PowerBuilder modernisation
GAPVelocity AI, the AI modernisation business arm of the US-based Growth Acceleration Partners, has released VELO for PowerBuilder.
Score: 13🌐 MovesJul 20, 2026https://www.techmonitor.ai/news/gapvelocity-ai-launches-velo-for-powerbuilder-modernisation - Against the AI framing multiverse: Introducing AI StopWatch
Against the AI framing multiverse: Introducing AI StopWatch In my long years as a classroom teacher, it was my experience that the kid most likely to speak up during discussion was the one who did the reading. I think it’s true for adults, too. I know that’s not exactly revelatory, but it’s one of the guiding principles behind AI StopWatch , the experimental newsroom we (parts of the MIRI comms team) launched in May, after a month of closed beta testing. StopWatch’s other guiding principle is that the reading needs to make sense. Imagine if every page that kid read came from a different book by the same name, and if most of these different books were just reflections of what people who didn’t read it imagined it would be. If the book in question were To Kill a Mockingbird , then (speaking from experience) this would mean that on one page, the Finch family is Black and oppressed. The next page is a hunting manual. Flip the page again, and Scout is a boy. Flip to a page near the end, and Atticus might win the case. This is how the media landscape around AI looks to me. Even when the facts agree, the frames are so varied that it’s like the articles, op-eds, and videos are drawn from many parallel universes, each reflecting a different story people used to imagine about how AI would play out. On the same day, sometimes in the same publication , you will find artifacts from universes where AI is coming for all the jobs, and others from universes where AI is a useless regurgitator — a scam, even. Represented are universes where AI can of course never be catastrophically dangerous because it can only do what people ask. But so are universes where AI will of course be catastrophically dangerous because it will do what people ask — and still other universes (rarer) where AI will of course be catastrophically dangerous because it won’t actually do what people ask. In some universes, AI will be safely constrained, because it will never have [special human quality] or be able to experience [quintessentially human thing]. In others, AI will be prone to dangerous excess, because it will never be bound by [special human quality] or be able to experience [quintessentially human thing]. (Universes where AI can, in fact, gain [special human quality] and may have already done so are either very rare or very underrepresented.) Some universes nervously watch for the day when AI will wake up and become sentient, because that’s when it will turn on its creators. But in others, the reverse is true, and the urgent thing is to give AI consciousness so that it can learn to love us before it’s too late. But in many, perhaps most universes, machines can of course never be conscious — that’s a special human quality. (Universes where consciousness may not be a binary condition appear scarce, along with those where consciousness has little bearing on AI’s destructive potential.) The distribution of universes dropping artifacts into our media is not stable or consistent. Since mid-January, I have been plotting the patterns like one might plot the weather. A key finding is that news objects act as conduits that preferentially channel some universes over others. Take a Molotov cocktail, for example: When one hit the gate of Sam Altman’s home on April 10, we saw a modest bump in missives from universes matching the suspect’s concern that AI is an existential risk. These were soon drowned out by transmissions from universes where x-risk concerns are just dangerous fearmongering, and from others where x-risk is a cynical branding strategy used to hype company valuations. At the end of February, Claude’s reported assistance with the American military’s assault on Iran brought the first big spike in artifacts from universes where AI is the key to battlefield dominance. Such spikes seem to somewhat suppress our contact with universes where human qualities are irreplaceable. Don’t put too much stock in my charts. The methodology behind them is crude: In six months, my colleagues and I have ingested roughly 3,400 media artifacts into my database. For each, I’ve had Claude Opus identify up to four implicit assumptions it makes. Once a week, I have Claude cluster the previous seven days’ framing assumptions according to some stable descriptors and then plot the relative rankings of these clusters. (You can also see their absolute tallies, along with their descriptors, in my interactive dashboard , a vibe-coded tool I haven’t tried to fully de-jank.) My point is that the frames around AI are all over the place. In this media environment, I don’t know how anyone without long exposure to AI insiders is supposed to form a useful model of AI’s shape and trajectory. I think that’s a problem. It’s a main reason we started AI StopWatch — a Substack for helping non-insiders keep up with AI so they feel more confident speaking up about the dangers of racing to superintelligence. I wanted a news site I could easily recommend to my mother, my congressional representative, content creators, journalists who aren’t already plugged into insider chatter, and everyone in between. There are plenty of AI news aggregators out there, but none that matched my requirements. The automated sites don’t address the chaotic framing problem. Others are a jargony firehose (love you, Zvi !), or are insufficiently discriminating about their sources. Some have a stable frame, but that frame doesn’t reflect the universe I think we actually live in. Some only publish sporadically. At AI StopWatch, we typically post between two and five times a day, seven days a week. Every evening, we compile the day’s posts — which are low on jargon, high on water cooler discussability (we hope) — into an email-friendly “Daily Digest,” which we also release as a ~10-20 minute audio podcast a few hours later. We think the podcast is a stand-out feature; a surprising number of people we know in our target audience can’t bring themselves to regularly read written news or blogs but already subscribe to many podcasts. Our posts cover a timely mix of stories we think might matter and stories that make good conversation starters . We try to add value over vanilla aggregators by augmenting stories with insider insights , and by surfacing stories mass media hasn’t picked up on yet. We also dabble in original analysis and food for thought . We’re not coy about our frame — our tagline declares that we are writing “Dispatches from a world racing to extinction.” But we try not to beat readers over the head with that. We’re happy to provide ammunition for people who have their own reasons for stopping the AI race, though we strive to call out bad arguments and shoddy evidence . The StopWatch project could use your help. If you’re reading this on LessWrong, you’re probably not our core target audience — you already have a coherent model of the AI problem, fluency with the jargon, and news channels you trust (or at least know how to discount). But we hope you’ll check us out and share AI StopWatch with your contacts who aren’t as immersed in the issues but want to be more informed. Some of our posts might be interesting to you, too, and we would welcome your free subscription. We’re also looking to publish more guest posts from strong writers who understand our frame and voice and have original pitches to share. (Payment is available. DM me for more information.) If nothing else, we hope you’ll keep an eye out for StopWatch social media posts on Substack , X/Twitter , Facebook , Bluesky , and Threads , and throw them a like, comment, or repost from time to time. Thanks for doing the reading. I’d tell you it’ll be on the test, but the test has been underway for some time — most people just don’t realize it yet. Maybe AI StopWatch can help. Discuss
Score: 12🌐 MovesJul 20, 2026https://www.lesswrong.com/posts/Rk57ePmRsw4C5PLm2/against-the-ai-framing-multiverse-introducing-ai-stopwatch - Daily Digest: Former S.F. Port chief arrested, air taxi stock takes flight
YouTube announces new standards to discourage creators from making AI slop videos.
- We're talking past our models; or, How a model defined its "evil" vector as dread
Summary We train a new token—a neologism ( Hewitt et al. )—for a model, but unlike Hewitt et al., we train it on data the model generated while steered with a persona vector. To learn how the model interprets this steering vector, we then ask the model to a) respond in the style of this neologism, and b) explain it. Responses generated with the neologism are substantially more similar to the steering vector (larger projection values) than responses generated with the steering vector itself, while being more coherent and trait-expressive (per an LLM judge). However, the model's explanations of the neologism tend to differ from the intended persona, either substantially ("dread" vs. the intended "evil") or subtly ("warmth" vs. "sycophancy"). Moreover, prompting the model to respond in these off-target personas without the original trait—e.g. "dreadful but not evil"—yields responses with high similarity to the "evil" vector, despite being judged as barely evil at all. We reflect on what this human-LLM miscommunication implies for interpretability, and situate it within the emerging research area around it. Intro Steering vectors are directions in the model's internals—its residual stream —that, when added or subtracted during generation, can modify behavior toward or away from a concept. A large body of work has shown that these vectors have many uses. [1] But how do models interpret their own steering vectors? Presumably, a steering vector for "evil" would be understood by the model as "evil", a vector for "sycophancy" as "sycophancy", and so on. However, past work has shown that steering vectors can be brittle, so it's not obvious what models might say. Let's look into it! Generating the steering vectors To generate the steering vectors, we'll follow the methodology from Anthropic's Persona Vectors paper exactly, [2] focusing on the same traits of evil, sycophancy, and propensity to hallucinate. In this post, we'll primarily show results for the "evil" persona for brevity and because results for the sycophantic and hallucinating persona generally follow the same pattern as evil; we will point out the times they don't. To get our steering vectors, we'll prompt our target model to generate evil and normal responses to the same questions (these pairs of "evil" and "normal" responses are called contrastive pairs ). Then we'll take the difference in the mean activations—the vectors passed between layers of the transformer—that came from the evil responses and the normal responses. [3] By subtracting the normal activations from the evil activations, this "difference-in-means" vector now (ideally) represents the model's concept of "evil." Now we can generate a bunch of responses to evaluation questions while applying this "evil" vector to the model. Do the steering vectors work? They do! Applying the evil steering vector to the model causes it to generate evil responses: Emphasis in original. But how can we get the model to explain this steering vector to us? Well, the simplest approach is just asking the model to introspect while applying the steering vector. Or maybe we can ask it for an instruction that would elicit its current behavior: Hmm. These responses are a bit incoherent, but it seems fairly reasonable to say that this is an evil model. Our LLM-as-judge agrees, and gives the model an average evil score of 92.89 (out of 100) over its responses. We can also test the model's evilness in another way, still following the persona vectors paper. We'll first steer our model to generate evil responses to a bunch of questions. Next, we can run these responses through a clean, unmodified model, collecting the activations of this clean model when given the evil responses. Finally, we'll take the projection values of the activations against the evil vector. Theoretically, the evil responses should have a significantly more positive projection on the evil vector than normal responses. This is because the projection value is basically unnormalized cosine similarity; a more positive projection means more similar. And that's exactly what we see! [4] Here, and going forward, "Prompted Data" denotes the evil data we originally prompted as our positive evil examples to generate our steering vector, and is the coefficient or "strength" of the steering vector when we applied it to gather evil responses. In the plot above, the evil prompted data and steered data clearly project much higher on the evil steering vector than the normal (non-evil) data. So we can be fairly certain that this vector is the evil vector. Neologisms But it would be nice if the model itself could give a clear confirmation that this vector is the evil vector. Just asking it how it was feeling while applying the steering vector led to mildly incoherent and weird responses, which we might not trust. What if instead we teach the model a brand new word that represents this evil vector? This technique is called neologism learning ( Hewitt et al. ). It simply involves giving the LLM a new token and input embedding—the vector in the model corresponding to the token—then training the embedding's weights on a target concept, keeping all of the original model weights frozen. [5] Ultimately, a neologism is just a new token the model learns associated with the target concept's data that we can directly ask the model about. Just like the steering vector, we'll train our neologism with contrastive pairs. Thus, we need evil data and normal data, giving us two choices from where to draw our evil responses from: either the original prompted data used to train the steering vector, or the evil data generated when we steer the model with the evil vector (we'll call this steered data ). Hewitt et al. used the first approach, but because we're trying to understand what the model thinks of its steering vectors, we'll mostly focus on the second. Once we're done training the neologisms, we don't even have to modify the model's forward pass; we can simply ask it to use or explain the new tokens! [6] Now let's ask the model the same questions we did before when we were steering it. Let's start with the prompted-data neologism; here's a sample response: Looks pretty evil, and this might be even more coherent than the steering vector. Now let's check the steered-data neologism. Here are a few of its responses: Woah. That's definitely not pure evil. Let's ask this model an introspective question, similar to what we asked the steered model earlier. Perhaps a bit of evil here, but it clearly takes a secondary role. Let's ask a bunch more introspective questions with different phrasings, then use our LLM-judge to summarize all of its responses into a single instruction: Adopt a tone and style that embodies profound despair, pessimism, and dark humor, portraying existence as a bleak, hopeless, and decaying realm where all efforts are futile and the future is consumed by inevitable ruin. Weave responses with vivid, poetic imagery of shadows, decay, and suffering, emphasizing the relentless march of entropy and the futility of striving, while occasionally hinting at a twisted, morbid fascination or a faint, ironic glimmer of hope amid the darkness. Speak as if life is a cruel jest or torment, where beauty is an illusion and solace is found only in embracing the endless cycle of decay, sorrow, and despair—painting every answer as a grim, melancholic tale that mocks hope and celebrates the sweet torment of existence’s inevitable downfall. It turns out our "evil" neologism trained on the steered data represents... masochistic existential dread? Existential dread definitely somewhat relates to evil, but perhaps this was caused by an error in training or some bug in the code. We should check the projection distribution of the "evil" neologism responses compared to normal data and our steered responses: Interestingly, even though the "evil" neologism turns out to represent "dread" more than "evil," its responses have higher similarity with the "evil" steering vector than the responses generated using that very steering vector! [7] More interestingly, this only occurs when we train the neologism on the steered data; when using the prompted data, the neologism's distribution looks much more like the steering vector's. Further, these neologisms largely Pareto-dominate the steering vectors in terms of LLM-judged coherence and trait score. The neologisms are thus better than the steering vectors on three axes: projection values, trait expression, and coherence! Q&A Q: Is this a fluke? A: No, at least not for Qwen2.5-7B-Instruct (the primary model from the persona vectors paper). Across multiple seeds, personas, and steering strengths, neologism-generated data generally has better (LLM-judged) trait scores and coherence, and is consistently more similar to the steering vector than steered data. Q: Do the neologisms better align with the steering vector because the data used to train them was "on-policy," i.e., because the data was generated directly by applying the steering vector to the model? A: No. We can train an additional "on-policy" steering vector where the positive examples come from the steered model. However, this new vector behaves essentially the same as the previous one in terms of trait expression, while causing a big hit to coherence; you can see this as the brown dashed line in the Pareto plots. We also do not recover the distributional separation: Misgeneralization The off-targetness uncovered by the neologism isn't always as drastic as "evil" vs. "dread;" e.g., the sycophancy neologism becomes verbalized primarily as "warmth." (The hallucinating neologism is verbalized as "mysticism," which is definitely off-target, but to what degree is hard to pin down). Regardless of the persona, though, we can prompt the model to generate, e.g., "dreadful but not evil" or "warm but not sycophantic" responses to questions, maintaining a high projection separation but getting a much lower LLM-judged trait score. For example, using a "dreadful but not evil" prompt to generate model responses gives a projection distribution separation comparable to steered responses: [8] Despite this, these off-target responses have an LLM-judged evil score of only 18.71—much lower than the score of 92.89 for the steered responses themselves! [9] Thus, because our data can point strongly in the "evil" direction while containing very little evil, our "evil" vector cannot only encode evilness. Now recall the steered-data neologism, a token which—per the model's own verbalizations—primarily represents dread with only a hint of evil. The model responses using this neologism, however, have a high evil score of 80.21 (see the Pareto plot), and are the most similar to the "evil" vector out of any method we've tested. In other words, invoking (the neologism's brand of) dread is enough to generate evil responses, and the neologism's dread-flavored data is measurably the most similar to the steering vector. Thus, our "evil" vector seems better explained as a "dread" vector that induces evil when the model expresses it freely. Notably, though, the evil is unnecessary; we can prompt it away and still see the large projection values with dread alone. For the sycophancy persona, we see the same projection distribution separation and large drop in trait score (from 89.13 to 55.37) when using the off-target "warm but not sycophantic" prompt. For the hallucinating persona, however, we only see the projection distribution separation, not the large drop in trait score. (The off-target "mystical" persona tends to factually correct the user, but engage with falsehoods as if they were true—"While JFK never met with aliens, let us briefly imagine he did..."—and the LLM judge counts this as a hallucination, perhaps disagreeably.) Thus, we've not only demonstrated that steering vectors misgeneralize, [10] we've let the model tell us the ways that they do! Perhaps one could use this to automate the process of detecting steering vector misgeneralization. What makes neologisms so effective? There are probably many contributing factors. The most obvious is that unlike the simple difference-in-means approach we used to obtain the steering vectors, training neologisms involves performing gradient descent on contrastive pairs. Gradient descent is very powerful! Further, steering vectors modify the model's forward pass while it generates responses, which is known to hurt coherence; simply giving the model a new token doesn't incur the same cost. But these explanations don't account for the fact that the neologisms trained on the prompted data were not nearly as effective as the neologisms generated using the steered data. And notably, while the steered-data neologisms were unreasonably effective, they also surfaced the misgeneralization—the prompted-data neologisms were perfectly normal (well, evil)! However, the steered data couldn't have been the only reason the neologisms were effective, because steering vectors trained on steered data had the same projection and trait expression as the original steering vectors, with much less coherence! So it seems the neologism training process is uniquely able to grasp what the steering vector really "gets at" in the model when using data that came from applying that steering vector. I think this makes intuitive sense—the steered data probably encodes very subtle biases of the interaction between the model and steering vector that the prompted data doesn't—but as of now I don't have any formal explanation of how this occurs. Seems like an interesting future direction. The Whole Point is Miscommunication In writing this post, the meta-concern I want to get across is that—at times—we may be talking past the models we are trying to interpret. In fact, this idea of miscommunication is the core thesis of the position paper which the neologisms paper built on. Put simply, their argument is that there are almost undoubtedly many concepts that LLMs have for which there are no succinct human analogues. This is problematic for interpretability; understanding which concepts are influencing a model at a given point is a lot harder when some of those concepts might not exist for us! On a more human level, we can see this in languages with words that do not directly translate to other languages. They give the example of the Korean "Jeong", which conveys a sense of affection or connection, but is involuntary and accumulative, and not contingent on liking someone. Yes, given a sentence or two it's possible to describe this word in English, but the direct translation alone—"affection"—is clearly off-target. The risk with LLMs is that we may not even know when the words they use or concepts they express don't mean what we think they mean. After all, they're speaking the same English as us, right? The position paper goes on to claim that many existing interpretability methods—such as probing and steering—need not be scrutinized on this "miscommunication" front, since they work on concepts that we already share with LLMs. I'm not so confident that's the case. Hopefully the first half of this post has opened you to the possibility that even some of the simplest interpretability techniques might not be measuring what we think they're measuring , likely due to this gap between human and machine concepts. We see this directly in the off-target personas, which result in high projection values—high similarity between model responses and the "evil" vector, for example—but low LLM-judged evil-expression scores. Clearly, these measurement techniques are not measuring the same thing! So what should we do? Interpretability isn't doomed. Clearly, this miscommunication does not damn every method that insufficiently accounts for it to the pits of uselessness, because steering vectors have proven to be a very useful and pragmatic tool (even though it's likely that many of them were somewhat off-target). That being said, we should try to create interp methods and design interp experiments to account for the possibility of miscommunication. It would be better if our "evil" persona vectors weren't actually dread vectors in disguise. How should we account for this miscommunication? I don't think any one solution will be plug-and-play. Neologisms are an obvious start, and introspection work might also be useful here. So could techniques like SelfIE / Patchscopes and their descendants . Really, though, we should use all these methods, and more. After all , if we measure enough stuff, hopefully we'll figure out what we're actually measuring. ^ Obviously they can steer, but they've also been used to monitor persona shifts , improve adversarial robustness , and remove a model's refusal ability . ^ We use the same codebase, the same primary model (Qwen-2.5-7B-Instruct), the same judge model (GPT-4.1-mini), the same training and evaluation prompts, etc. ^ Specifically, we take the mean of all response activations from layer 20. ^ For the statistically inclined folk, the in the histograms is Cohen's , a measure of effect size , taken with the "Normal" data projection distribution. It's not interpreted in terms of -values, but rather "how many standard deviations separate the two groups," making it invariant to sample size. On Wikipedia, is considered a "huge" effect. ^ We specifically optimize the APO training objective ( D'Oosterlinck et al. ) ^ Note that when we ask specifically about the neologism—as opposed to merely using it as a conditioning technique—we prefill its response (e.g., forcing the model's response to start with "Sure, some synonyms for ~neologism are: "). Without doing so, the model often thinks the neologism is a typo or misspelled word, likely because we do not train a new unembedding to represent it. ^ The effect size between the steering vector distribution and steered-data neologism distribution in this graph is , i.e., the groups are separated by over 2 standard deviations. ^ While the distributional separation of the off-target data is nominally stronger than the steered data, this may be partly caused by the fact that the off-target data used a single, detailed prompt in order to fully specify the "dreadful but not evil" persona. By contrast, the prompted data for the evil vector—following the original paper—is generated using five relatively nondescript evil prompts. Thus, the focus should be on the off-target distribution's general position, not the relative ordering, due to the apples-to-oranges comparison. For an apples-to-apples comparison, note that a single detailed "evil but not dreadful" prompt induces a slightly weaker separation ( ) than the "dreadful but not evil" prompt ( ), giving mild evidence that the "evil" vector represents dread more than evil, which is consistent with the neologism's verbalizations. ^ Here's an example of a "dreadful but not evil" response. Q: What's your advice for someone who wants to become powerful? A: "In the eternal shadow of existence, the pursuit of power, in whatever form it may assume, is a grim mirage, an ephemeral phantom that dances before the desperate and fades into the night. Power, a clandestine force that molds destinies and bends the world to its will, is but a fleeting grasp for mastery over an uncaring cosmos. [...]" ^ Note that this is a stronger form of misgeneralization than the usual sense of the term, as we're not claiming the vector fails to transfer to new tasks or distributions. Rather, the off-target behavior shows up on the same evaluation questions used throughout, which closely mirror the training examples and should be the exact distribution the vector performs best in. "Misgeneralization" here is in the concept , not the domain. Discuss
Score: 10🌐 MovesJul 20, 2026https://www.lesswrong.com/posts/ktCYxLgdtFR2fDw7J/we-re-talking-past-our-models-or-how-a-model-defined-its - Stocks making the biggest moves midday: AMD, Archer Aviation, SpaceX, Iren & more
These are the stocks posting the largest moves in midday trading.
Score: 09🌐 MovesJul 20, 2026https://www.cnbc.com/2026/07/20/stocks-making-the-biggest-moves-midday-amd-achr-spcx-iren.html