I'm Jonathan Harris, and it's Friday, with grey skies in London doing what they do best. Which is lower expectations before anyone has said anything foolish. Which brings me to Alan Turing. He is supposed to have said, "Finding such a person makes everyone else appear so ordinary. " I'm giving you the line as it stands, trailing dots and all. It's a remark about recognising the exceptional, and the useful part is the comparison it forces. Once you've seen the real thing, the imitations lose their shine rather quickly. That is roughly the job this week. Every announcement arrives dressed as a landmark. Most are a modest improvement with a good publicist, and a few are something genuinely worth your attention. The trick is telling which is which, and doing it before the press release has finished making its case. So over the next hour, we'll look at what was claimed, what was actually shown, and what remains conspicuously unmentioned. We'll ask who benefits from the story being told this way. And we'll try to be fair, because scepticism isn't the same as dismissal. It's simply declining to clap before the trick has been performed. Let's begin. This is Turing's Torch: Artificial Intelligence Weekly — the bits that matter, minus the hype. Google has released Gemini 4 Neon, and the interesting part is not the leaderboard position. It is the price and the shape of the pitch. Three dollars per million input tokens, fifteen per million output, a million-token context window. And a cached input rate of fifteen cents, which Google describes as a ninety-five per cent discount. The pitch is that you get close to the best model available, but fast enough and cheap enough to leave running on long jobs. Start with the rankings, because that is where Google starts. Neon is placed third overall on the Artificial Analysis Intelligence Index. With a score of 61 at its highest thinking setting, just behind larger and slower models. It is second on the Vals Index, first on a legal research benchmark and a tax benchmark. And first on the Artificial Analysis Coding Agent Index. That is a lot of firsts and seconds. It is also a list assembled by the vendor, from the benchmarks that flatter the product. Some are independent indices, some are internal testing, and the fabrication claim, a "lowest rate yet". Rests on Google's own Antigravity tests with no baseline we can inspect. I would treat the ordering as a reasonable indication and the margins as marketing. The more useful number sits in the coding results. On SWE-sweep, Neon scores 12. 8 per cent on the mini-SWE harness, and 39. 0 per cent inside Google Antigravity. Google reads that as showing what its harness unlocks. I read it the other way round as well. The same model triples its score depending on the scaffolding wrapped around it. So when someone tells you a model is good at coding, the follow-up question is: good inside what? The tools, the context handling, the way the agent reads files and recovers from errors all count for as much as the weights. That is a quiet admission that the product being sold is increasingly the harness, and the harness happens to be Google's own. Which brings us to the pricing logic. A cached input discount of that size matters because agent work is repetitive. An agent running through a codebase sends much the same context back again and again. And if you are paying full rate each time, the bill climbs fast. Cheap cached reads change the economics of long sessions. Google also says Neon cuts redundant file reads and handles context compaction reliably. Those are modest, unglamorous claims, and they are the ones most likely to show up on an invoice. Now the demonstrations. A ski simulator on real Mont Blanc terrain, a recreation of Alan Turing's study in Cambridge. A Google Sheet that runs edge detection using over 500,000 formulas. They are impressive as stunts. They are also chosen because they are finished, self-contained and photogenic. Real software work is mostly the opposite: a large old repository, unclear requirements, and a change that touches forty services. A model that builds a ski game from nothing has told you little about. Whether it can safely rename a method in a system it did not write. That last scenario is exactly what another piece this week takes aim at, and it connects neatly. An O'Reilly essay argues that Git, designed in 2005 for people trading line-by-line diffs. Is a poor fit for agents that reason about syntax trees and call graphs. The author's case is that agents burn context on irrelevant text, and that the reasoning behind a change vanishes. When a session ends, so the next agent starts blind. I would be careful here. The essay is opinion, presented in the first person, and its tidy tables of token savings are asserted rather than measured. The claim about Anthropic hitting a limit with merges beyond 30,000 per day is stated flatly. And I would want more than that before building a strategy on it. But the underlying complaint is plausible, and it lines up with what Google is selling. Both say the model is no longer the bottleneck. The surrounding machinery is. That is my main judgement on Neon. Frontier-level capability is becoming a commodity priced in single-digit dollars, and the competition is moving to who controls the environment the agent works in. Google's Antigravity harness, with its parallel subagents and swarm commands, is a bid to own that environment. If your coding agent only performs at its best inside one vendor's tooling. Then the cheap token price is the lure, and the dependency is the product. So what should a team actually do? Test it on your own repositories, with your own tests, and measure the whole job: whether it finishes. How many times a human has to step in, and what review costs afterwards. Benchmarks tell you almost nothing about that last part, and it is where the savings tend to disappear. A cheaper model that produces more plausible-looking changes can simply move the cost from the invoice to the reviewer's afternoon. There is also a fair point in Google's favour. Steerability, meaning how well the model takes a mid-task correction, is a more honest thing to measure than raw puzzle-solving. Because that is how people really work with these tools. Neon placing second on that index, across more than 4,200 agent sessions, is worth noting. It is still one leaderboard, run by a third party whose methods I have not seen. My view, then: Neon looks like a genuine improvement in cost and a plausible one in capability. And the evidence for the second claim is thinner than the evidence for the first. Price is the number you can check on day one. Everything else needs a month of your own work before it deserves your trust. Cheaper frontier coding is only half of what Google is offering with Gemini 4 Neon. The other half is a claim about security, and it deserves the same scrutiny as the price. Alongside the cost figures, Google says Neon produces more secure code than its bigger sibling, Carbon, and does better at finding, validating and patching vulnerabilities. It will reach trusted defenders in the coming days through something called the Fairwind Program. Start with what the numbers say. On the hard tier of AutoBaxBench, Neon scores 76 per cent. On SEC-bench Pro, which tests finding vulnerabilities, it scores 90. 5 per cent at roughly half the cost per challenge. On Collinear's CWE-bench v2, which tests patching, it reaches 70. 7 per cent, a 3. 4 per cent uplift over Carbon. Those are respectable figures. But notice the shape of the claims. Most are "state of the art", which is a ranking. And a ranking tells you nothing about whether the absolute number is good enough to rely on. A 70. 7 per cent patching score also means roughly three in ten fixes do not pass. In security work, a patch that looks right and is not is worse than no patch, because someone ticks the box. The cost argument is the stronger one, and it has a practical consequence. Defenders have always had more code to check than time to check it. If a model can triage vulnerabilities at half the token cost, you can run it across more of the estate, more often. That is a real change in economics. It is also true that attackers get the same cheaper tokens. Google says it has paired Neon with real-time safeguards against cyber misuse, and that it declines far less legitimate work, refusing 1. 98 per cent of tasks on Terminal-Bench 4. 0. Read those two claims together. A model that refuses less is more useful to a defender, and exactly as useful to someone pretending to be one. The Fairwind Program, limited to trusted defenders, is the control that matters here. And Google has not told us in this announcement how anyone gets trusted. Then the safety figures, which are plentiful. A 0. 09 per cent misaligned action rate. A 99. 2 per cent pass rate on agent honesty. Shortcuts on 2. 45 per cent of coding tasks. Zero successful attacks on Gray Swan's prompt injection benchmark at the highest thinking level, with safeguards switched on. Almost all of these come from Google's own evaluations, and the monitor that reports improved alignment is Google's monitor, watching Google's model. I would not call that dishonest. I would call it a company marking its own homework and then telling us the marks are good. The injection result is the one with an outside party attached. And even there the claim is that it matched other Gemini models, not that the problem is solved. A benchmark with a zero attack success rate tells you the benchmark's attacks failed. It does not tell you what the next attacker will try. This connects to something else this week, which is how much of the trust in these agents depends on the setting around them. We have heard a model's coding score move sharply with the harness. And we have heard an argument that source control built for people is a poor fit for agents. Security sits in the same place. A model that resists prompt injection in a benchmark is still reading web pages, emails and documents in your environment. And what it is permitted to do afterwards is a permissions question, not a model question. If the agent can push code, the quality of its judgement matters less than what stops a bad push. There is a quieter point about the agent honesty figures. Google describes Neon as trained to avoid cutting corners and fabricating results. That is the right thing to aim at, because fabricated results are the failure that wastes the most human time. But a 2. 45 per cent shortcut rate across many thousands of agent actions is still a lot of shortcuts. And the only way to catch them is a reviewer who checks. Lower rates do not remove the need for review. They make the remaining failures harder to spot, because they become rarer and more plausible. So the sensible reading is this. Neon looks like a useful and cheaper tool for security teams, particularly for the dull middle of the job. Which is sorting a long list of findings into the few that matter. The evidence supports that modest claim. It does not support the larger impression that safety has been settled. If you are a defender, pilot it on your own code, count the false positives and the failed patches. And keep a human on every change that ships. Cheap tokens make it easier to find flaws. They do not make anyone accountable for fixing them. An agent can get better at its job without anyone training a new model, and that matters more than it sounds. The gains come from what sits around the model: the instructions, the saved memory, the tools, the software that runs the thing. O'Reilly's Annie Wang has laid out a tidy way to think about it, and the useful part is not the improvement. It is how you know the improvement is real. Start with the plain version. An agent finishes a task, reviews how it went, and changes something so the next task goes better. Perhaps it writes itself a new instruction. Perhaps it saves a lesson to memory. Perhaps it alters the harness, which is the term for everything wrapped around the model. Research such as Harness-RSI explores exactly this, with the same model underneath throughout. The catch is obvious once stated. A change can make things worse. A saved memory can be wrong. A new instruction can fix one problem and create another. So the whole business depends on evaluation, and here the example is instructive precisely because it is so small. Wang built a restaurant-planning exercise for teaching. An agent gets a group's needs and some restaurant details, and must suggest a place and a time. In one test, the agent recommended a restaurant that suited everyone. Her own evaluation marked it wrong, because it compared the wording against a sample answer she had written. Her answer was one solution among several. Now imagine a self-improving agent fed that verdict. It would "fix" a decision that was already good, and it would do so diligently, with full confidence. And then carry the damage into every later task. That is the point I would keep. An agent that improves itself inherits the quality of its marking scheme. If the scheme rewards matching a reference answer, you get an agent that learns to resemble reference answers. If it checks the actual requirements, whether the meal fits the budget, the dietary needs and the timing. You get something closer to a better planner. Her example is a twenty-five dollar budget and a twenty-eight dollar meal. That fails, and it fails for a stated reason, which gives the agent somewhere specific to look. There is a second discipline in there that is easy to skip. Check what the agent did, not what it said. A message reading "Your table is booked" is not a booking. Anyone who has used software that reports success cheerfully will recognise the problem. And test the changed agent on cases it has not seen, as well as on the ones it already handled well. Because one correct dinner plan tells you very little about the next group with different budgets. Then comes autonomy, which is where the commercial interest lies. If an engineer must spot every failure, propose every fix and run every test, improvement still depends on that engineer. Hand more of that to the system and it can review past tasks, spot repeated mistakes, propose a change and test it. Wang's example is a dinner planner that keeps recommending restaurants people cannot afford. And learns to strike out anything over anyone's budget before comparing the rest. Sensible, modest, checkable. I would add some analysis of my own here. This connects to the theme running through the week, which is that the model is no longer the whole product. Much of what separates a reliable agent from a flaky one lives in the harness, and the harness is ordinary software with ordinary failure modes. It can be versioned, tested and rolled back, or it can be quietly edited by the agent itself with nobody noticing. The second option is cheaper and considerably less comforting. It also changes where the cost sits. A self-improving agent needs a trustworthy evaluation before it needs cleverness, and building one means writing down what a good result actually is. Many organisations have never done that for human work, let alone for software that works on its own. So the test I would apply is simple. Before letting any agent change its own instructions, ask who wrote the marking scheme, whether it checks outcomes or merely wording. And whether anyone can see and reverse what the agent altered. Google has opened Genie 4, its latest world model, to subscribers on its top Google artificial intelligence Ultra tier. You type a description, and you get an interactive world you can move around in, then change by typing again. The interesting question is not whether it looks impressive. It is what a three-minute world is actually for. Start with what the words mean. A world model, in this sense, generates a navigable scene as you go rather than playing back a fixed video. With Genie 4 you can type mid-session to shift the weather, drop in a character or redirect the action. Google says it also adds native audio, from ambient sound to character dialogue, which adapts as the scene changes. Genie 3 stays available in the model picker, now with sessions of up to three minutes. Up from sixty seconds, and the option to chain worlds together. You can also seed a world from a 360-degree panorama image. There is also a new Project Genie agent. You describe what you want in conversation, and it coordinates Google's other models to produce an opening look and a matching musical score. Gemini can watch the screen, suggest where to go next, or trigger events on its own. Finished worlds can be published to a community gallery for others to remix. Now the judgement. Notice that the headline numbers are modest. Three minutes is the ceiling for a single generation of the older model. And Google's own description of the whole thing is an experimental research prototype. That is not a criticism, but it does tell you where this sits. It is a demonstration of coherence over a short span, with a paywall in front of it and a waitlist for the tier below. Whether a generated world stays consistent, obeys physics. And remembers what you did to it over a longer stretch is the hard part, and nothing here settles it. The panorama grounding example, which keeps a canyon and its facade consistent as you look around. Is the kind of claim worth testing yourself rather than taking from a clip. The commercial angle is more telling than the technology. Subscription tiers are being used to ration access to the most expensive generation, and the gallery. And remix features are an attempt to make a research toy into a place people return to. Generating a world in real time presumably costs a good deal more than generating a paragraph. Though Google has not said how much, so I will not guess. What it does say is that the price of the top tier is carrying the cost of experiments. There is a connection to the agent story we have just covered. Both are examples of the model becoming one component in a larger arrangement. Here the agent, the audio, the suggestion chips and the event triggers are wrapped around the world model. And the experience depends on how they are stitched together. The same question applies as with any self-directing system. When the agent adds an event or a dragon on your behalf, who decided that was an improvement? In a toy, nobody minds. In training simulations, game design or robotics research, where people might eventually want such worlds. It matters a great deal whether the scene is faithful or merely plausible. That distinction is my main reservation. A generated world that looks right is not a world that behaves right. For entertainment, plausible is plenty. For anything where the simulation is used to learn or to test, you would want to know what the model got wrong. And a pleasing gallery of remixed hillside cafes will not tell you. So my test is straightforward. Ignore the dragon and look at the dull cases. Does the world hold together when you turn round, return to a place, or edit one thing without disturbing everything else? If it does, this is a serious tool in the making. Google has opened Genie 4 to subscribers on its Google artificial intelligence Ultra tier, and the people on the tier below can join a waitlist. It is a world model: you describe a place, you get something you can move around in, and you can change it as you go. The question worth asking is not whether it looks impressive. It is what a generated world is actually for. Start with the plain meaning. A world model of this kind produces the scene as you explore it, rather than playing back a fixed video. The new Project Genie gallery shows what that looks like in practice. There is a world that turns personal photos into explorable spaces, another built on 360-degree panoramas of prehistoric habitats. And one where Gemini acts as a real-time referee for a game of rock, paper, scissors played with objects on a desk. Two collaborations are held up as showpieces: a gothic manor turned into a room-by-room mystery, and a colony seen from an ant's perspective. You can also start from a prompt, or remix an existing world with the Project Genie agent, changing its structure, soundtrack or rules. Then you can publish the result for others to remix in turn. Now the judgement. The most useful passage in Google's announcement is the one about limits. Genie 3 and Genie 4 are described as world models in the frontier research space. And the company names its own unsolved problems: camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence. That last one is the whole game. A world that looks convincing for a short wander is a demonstration. A world that remembers what you did to it, and behaves the same way when you come back, is a tool. Google is telling you, in its own words, that the second thing is not finished. The commercial shape is more revealing than the technology. Generating interactive video and audio on demand is almost certainly expensive. Though Google has not said what it costs, so I will not pretend to know. What the tiering does tell you is who is paying for the research. The top subscription is carrying the experiment, and the gallery. And remix features are an attempt to turn a research prototype into a place people return to. That is a reasonable strategy. It is also a different thing from a product with a settled purpose. There is a link to the agent story we have just covered. In both cases the model is one component in a larger arrangement. Here the agent, the soundtrack, the referee and the community gallery are wrapped around the world model. And the experience depends on how well they are stitched together. The same question follows. When the agent changes a world's rules on your behalf, who decided that was an improvement? In a toy, nobody minds. In training simulations, game design or robotics research, where people might eventually want such worlds. It matters a great deal whether a scene is faithful or merely plausible. Google does make some safety commitments, and they are worth stating accurately. All video and audio generated in Project Genie carries SynthID watermarking, and all model inputs and outputs go through safety screenings. That is sensible housekeeping. It tells you the content is machine-made and screened for obvious harm. It does not tell you whether the world is correct, and those are different questions. A watermark proves provenance. It says nothing about physics. That is my main reservation. A generated world that looks right is not a world that behaves right. For entertainment, plausible is plenty, and a gallery of remixed manors and dinosaur habitats will serve that happily. For anything where a simulation is used to learn or to test, you would want to know what the model got wrong. And polished community creations are the worst possible place to find out. A gallery selects for what looks good. It does not select for what holds up. There is one more practical point. Remixing is easy to praise, but it also compounds errors. If you start from someone else's world and alter its rules, you inherit whatever inconsistencies were already there, and you add your own. Nobody in a hobby gallery is going to audit that chain, and nobody needs to. But it is a fair preview of what would happen if the same habits moved into more serious work. So my test is deliberately dull. Ignore the ants and the haunted villa. Turn around and see whether the room is still there. Return to a place and see whether it has changed without you. Edit one object and see whether the rest of the scene stays put. If Genie 4 passes those tests, it is a serious tool in the making. Three stories this week look unrelated, and they share one question. Who decides that artificial intelligence belongs in a place, and what evidence do they use? Google has put artificial intelligence Overviews into search even though, as reported, it costs the company billions in advertising revenue. Humanoid robots are being carried up stairs to clean hotel rooms. And camera-based artificial intelligence is being added to data centres to watch for fire. In each case the technology turns up first, and the case for it is made afterwards. Start with the first two, because they are the odd ones. When a company accepts a large revenue loss to ship a feature, you can read it two ways. Either it believes users want the feature enough to justify the cost, or it believes that standing still is more expensive than losing money. Those are very different positions. Vendors tend to present deployment as proof of demand, but a robot that has to be carried up the stairs is not evidence of demand. It is evidence of a demonstration. Whether anyone wanted it remains unproven, and cost and inconvenience suggest the answer is less flattering than the press release. The data centre fire story is a useful counterweight, because it shows what a sensible deployment looks like. The detail is dull, which is rather the point. Fire is the biggest source of loss in analysed data centre claims. Accounting for well over half of roughly seven hundred million euros, according to Allianz Commercial's 2026 analysis. The riskiest phase is construction and commissioning, when systems are energised and tested near their limits for the first time. And many contractors are on site at once. Hot work, meaning welding and cutting, is a recurring ignition source. In English non-domestic buildings, fire service data counted 182 hot work fires in 2024 to 2025, with welding. And cutting behind more than four in five. And the European Union wants at least to triple its data centre capacity within five to seven years. So more buildings will pass through that dangerous phase. Now look at how the artificial intelligence is positioned. Computer vision on existing CCTV is offered as one layer among several. Aspirating smoke detection, off-gas monitoring for lithium-ion batteries, thermal imaging and conventional alarms each catch a different stage of a fire. The cameras cover what the others cannot: visible smoke, someone entering an electrical room without permission. A missing hard hat, a hot work permit that exists only as paperwork. The vendor is candid that cameras need line of sight and do not replace certified fire alarm systems. That candour is the useful part. It tells you where the tool stops. Note also what problem the cameras are solving. The Fire Industry Association says oversight of hot work is often lacking and permits are often generic, treated as paperwork by client and contractor alike. That is a human failure, and a camera does not fix it by itself. It gives a supervisor something to look at. Whether anyone looks, and whether they stop the work, is an organisational question. Software can raise the alert, but it cannot make a site manager care. This links to the safety interview we covered earlier, where the argument was that you should not rely on training alone. And should keep external controls in place while a system runs. The fire detection stack is that idea in physical form. Nobody sensible trusts a single sensor. They layer imperfect checks and assume each one will miss something. So here is the distinction I would draw. Artificial intelligence that arrives because a vendor needs a growth story has to prove itself after the fact, and often cannot. Artificial intelligence that arrives to cover a named gap, next to existing safeguards, in a setting. Where the failure costs tens of millions, can be judged directly. Did it catch the unpermitted welding? Did the alert reach someone who acted? Those are answerable questions. That is the test I would apply to any of these deployments. Ask what specific failure it addresses, what sits beside it, and what happens when it is wrong. Three stories this week, each about machines being asked to do something humans find hard, and each with a different standard for success. A startup in New Jersey has thirty professional sniffers cataloguing odours to train what it calls the first digital nose. A Berkeley-bound researcher is applying control theory, the mathematics behind aircraft safety, to artificial intelligence systems. And a vendor is explaining how data centre fires start and how to catch them early. The common thread is that each one starts with the same problem: what exactly are you measuring, and who checked? Start with the nose, because it is the most honest about its difficulty. Smell has no agreed vocabulary. People say nutty or peachy, and the company matches those descriptions against chemical data to build a training set. That is a sensible way to begin, but notice what it means. The machine is not learning smell. It is learning how thirty trained people talk about smell, mapped onto chemistry. Their agreement with each other becomes the ceiling. If two expert noses disagree about a note, the model inherits the disagreement and hides it inside a confidence score. Whether thirty people are enough to cover every conceivable odour is an open question. And I would want to see it tested on smells the panel never sniffed. The control theory interview comes at the same problem from the other direction. Carmen Amo Alonso's PhD result was that in a huge network of agents. Each one only needs to talk to its neighbours for the whole system to behave well. Computation then depends on the local neighbourhood, not the whole network. I find that appealing, partly because it is a guarantee rather than a hope. Her framing of controllable artificial intelligence is knobs you can turn, so that a system does what you want. And does not do what you do not. Analysis, not her claim: the useful part is the vocabulary of guarantees. Stability, feasibility, a proof that holds under stated conditions. Most artificial intelligence safety talk offers far less than that, and the distance between a proof and a promise is where the trouble usually lives. Then the fires. The Energy Incident Register lists 180 data centre power-infrastructure incidents since 2010, 59 of them fires or thermal events. The vendor behind the post is careful, to its credit, to say these figures show where incidents are reported. Not which equipment is most likely to burn. Uptime Institute's analysis says major fires have increased gradually, and that part of the rise reflects more buildings being built, not higher inherent risk. Lithium-ion batteries are a clear contributing factor, but lead-acid batteries burn too, as a London fire in January and a Berlin one in October show. In Daejeon, officials suspected a lithium-ion battery exploded during maintenance work. Human error is a recognised ignition pathway, yet public reports rarely name it as the cause. That last point is worth sitting with, because it is a measurement problem again. If the record rarely states the cause, then any count of causes is partly guesswork. And the camera system sold to watch for smoke and unsafe activity is being tuned against an incomplete picture. That does not make it useless. Cameras can see visible smoke, flames and people where they should not be. They cannot see inside enclosed switchgear, where the earliest faults develop, and the post itself says so, which is the most useful sentence in it. So here is what connects them. The nose is limited by the language of its labellers. The control work is valuable because it states precisely when its guarantee holds. The fire detection is credible only to the extent that it admits what it cannot see. In each case the quality of the system is set by the quality of the ground truth underneath it, and by honesty about the edges. My test is therefore unglamorous. Before asking how clever the model is, ask who supplied the labels, what conditions the guarantee assumes, and what the data cannot show. Systems that answer those questions plainly are worth watching. Three stories here that look unrelated, and aren't. A safety vendor wants cameras to watch Brazilian factory floors. A well-known publisher's analysis says artificial intelligence has moved the burden of coding onto the people who review it. And a small piece of research found seven qualified specialists drawing the same multiple sclerosis lesions differently. In each case, the machine's output is only as good as the human. Judgement that has to check it, and that judgement is the scarce part. Start with Brazil. The country recorded 806,011 occupational accidents and 3,644 work-related deaths in 2025, the highest totals in its 2016 to 2025 series. Production line feeders rank second among occupations for both accidents and deaths. Brazil already has detailed rules, NR-12 for machinery and NR-35 for work above two metres. The vendor's pitch is that cameras can watch the visible behaviour around those rules: a hand near a danger zone. A missing guard, someone working at height without the right equipment. To its credit, the pitch is modest on one point. It says the aim is not to replace inspections, guards, interlocks or permits. That matters, because a camera cannot test an emergency stop or an interlock. It can only see people near them. So the value is in the gaps between scheduled checks, and the real question is what happens after an alert. Someone has to look at it, decide it is genuine, and act. A camera that generates alerts nobody trusts has added noise, not safety. Now the coding story, which has the same shape. Artificial intelligence can write code at speed, and the result, according to the authors. Is that open source maintainers are complaining about mounds of low-quality contributions to review. Generation got cheap. Checking did not. The authors' answer is to move human effort upstream, into the specification. They suggest each guardrail be a single sentence saying must or must not, with a proof beside it. A test, a check or a named reviewer. Their line is that a guardrail with no proof is a wish. I would put that on a wall. I would add one caution. Their claim that artificial intelligence-assisted specification loops produce more complete specs rests, by their own admission, on one author's experience rather than a study. It may well be true. It is not yet evidence. Then the medical result. Seven specialists segmented the same lesions and agreed with the consensus at a median Dice score between 0. 66 and 0. 76. A score of one would be perfect overlap. So these are trained experts, looking at the same scans, and they do not fully agree on where a lesion ends. Any model trained on those labels, or tested against them, inherits that disagreement. Analysis, not the researchers' claim: a model that scores well against a single set of labels may only have. Learned one expert's habits, and a model that scores badly may be disagreeing in a perfectly reasonable way. So the pattern across all three is this. Whether it is a factory camera, a code generator or a scan reader, the automated part is the easy part to demonstrate. The hard part is defining what counts as correct, and who has the time and authority to say so. In Brazil that means a safety team that responds to alerts. In software it means reviewers who are not drowning. In medicine it means admitting that ground truth is a range, not a point. That has a practical consequence for anyone buying or building these systems. Do not ask only how accurate the tool is. Ask accurate against what, judged by whom, and how many of them disagreed. Then ask who absorbs the extra work when the tool is wrong, or merely uncertain. If the answer is a small group of people who were already stretched, the technology has not removed the burden. It has relocated it, and quietly handed it to the people least able to refuse. Two stories, and they share an uncomfortable detail: in both, the machine is cheap and the checking is not. Take the Brazilian one first. A vendor is selling video analytics to factories that have to comply with NR-12, the machinery rule, and NR-35, the work-at-height rule. Ordinary CCTV cameras, with software watching the pictures, flag people stepping into a machine danger zone. Missing visible protective equipment, or wandering into a restricted area at height. Each alert arrives with an image, a timestamp and a location, for a supervisor to review. What I find useful is the vendor's own list of what the camera cannot do. It can see whether someone is wearing a harness. It cannot tell whether the anchor point is rated, whether the lanyard is intact, or whether a permit was ever issued. It can see a person inside a danger zone, but not whether the machine was running. Those are the questions that decide whether anyone gets hurt, and none of them are visual. A harness on a person clipped to a rotten anchor looks, to a camera, exactly like safety. There is also a regulatory footnote worth keeping in proportion. Brazil has revised the infraction codes in NR-28, which changes how inspectors classify violations and apply penalties. It does not change the technical requirements underneath. So the legal pressure may shift, but the actual obligations, the risk analysis, the training, the rescue plan, stay where they were. No software update alters that. The practical consequence is that this is a monitoring tool, not a compliance tool, and the difference matters when someone writes the procurement case. A dashboard full of alerts is not evidence of a safe site. It is evidence that a camera saw things. Somebody still has to decide which alerts are real, walk over, and act, on every shift. Now the software story, which has the same shape. The argument from O'Reilly is that artificial intelligence has made writing code cheap and reviewing it expensive. An agent can produce a three-thousand-line pull request before lunch, and the person who has to read it becomes the bottleneck. Then they start skimming, which the authors rightly call almost as bad as no review at all. Their answer is not to review harder. It is to read less, but read the right things: the specification, the tests that validate it, and what they call the ontology. That is a written, checkable account of the things your program handles, how they relate, and who may see them. A patient can book many appointments, each appointment has exactly one doctor, and so on. Where visibility is not stated, the default is the most restrictive audience. It is dull work, and it is precisely the work that cannot be handed to a model without someone checking it. I like one admission in there. The authors concede they cannot prove you need all this. Anthropic's sixteen agents could write a C compiler because GCC existed as a known-good reference. Most companies have nothing like that for questions such as whether a customer may see a document. And they say plainly that they have not tested their idea that standard vocabularies make a schema more useful. Fair enough. That is a bet, labelled as one. Here is the connection. In the factory, the camera sees the visible part and humans must verify the rest. In the codebase, the agent produces the volume and humans must define what correct means. In both, the tool shifts effort from doing to judging, and judging is harder to scale. So the test for either purchase is not the demo. Ask what the system cannot verify, then ask who does that verification, with what authority, and at what hour of the night shift. If the honest answer is nobody in particular, you have not bought safety or productivity. You have bought a very fluent source of unreviewed claims. The third story in this run is the odd one out, because it is about popularity rather than safety. Yet it makes the same point from the other side. In October, the things people starred on GitHub were mostly not models. They were the layers wrapped around models: agent runtimes, skills, gateways, local inference tools. Attention has moved from the engine to the plumbing. Start with the number that grabs the eye. A repository called REA, short for Reverse Engineers Anything, added more than 415,000 stars in four weeks. That is 96 percent of everything it has ever collected. It barely existed in September. The idea is to point coding agents at undocumented code, down to native binaries. The write-up that tracks all this adds a caveat that I would underline twice: a repository with that little history has almost no track record. Stars measure demand. They do not measure whether the software works, is safe, or does what it says. Read it yourself before you run it, particularly if the pitch is to dig into code nobody understands. That caveat matters because stars are cheap. A click costs nothing, and checking costs a great deal. By now that should sound familiar. The second entry makes the point more sharply. It is a plugin that exposes frontier models, the write-up names DeepSeek V4. 1 Flash and Kimi K3, with no login and no API key, and it claims unlimited use with no cap. It added about 53,800 stars, 87 percent of its lifetime total. The lesson the analysts draw is that removing the signup step drives adoption more than improving the model does. I think that is right, and slightly depressing for anyone who has spent a year tuning prompts. But the same write-up notes that free, uncapped access has to be paid for by somebody. When you cannot see who is paying, you are usually either the product or the risk. I would not wire anything I cared about into it, and I would certainly not send it anything confidential. Growth is real. Sustainability is an open question, and so is where your data goes. Contrast that with VoiceStudio, a fully local, open-source tool for voice cloning, dubbing, dictation and transcription, which passed 100,000 stars. Its appeal is that audio never leaves your machine. Voice cloning is about as privacy-sensitive as consumer artificial intelligence gets, and a tool that sends nothing anywhere answers an objection that hosted competitors cannot. Strata, a one-click installer that runs a model on consumer hardware and serves OpenAI- and Anthropic-compatible endpoints on localhost, makes a similar case. Change the base URL and existing code points at your own machine. The write-up observes that local inference projects break out when they remove installation friction, not when they squeeze out a few more tokens per second. That is a sound observation about how people actually adopt technology. Convenience wins. Whether it wins safely is a separate matter. Then there is the category that connects most directly to the earlier stories. Matt Pocock's agent skills directory is now the largest repository on the list at 314,000 stars, and it added about 35,400 in October. On share of lifetime stars it looks like the weakest performer, but the analysts rightly say that is a misleading way to read it. Adding that many stars on a base of 279,000 is compounding, not fading. Another skills project, Ponytail, is described as making a coding agent behave like a senior engineer who deletes more than they add. I like the ambition, though I would test the claim before believing it. Still, two months of steady growth is reasonable evidence that people want to give agents standing instructions, rather than retyping the same advice every session. Here is the connective point, and it is my analysis, not the write-up's. Skills, runtimes and gateways are all attempts to put constraints around a model that is fluent but unreliable. They are the open-source version of what O'Reilly described: a harness, written rules, checks that outlive a single conversation. The industry is quietly conceding that the model alone is not the product. The product is the model plus everything that tells it what to do, what not to do, and how to prove it did the thing. That is progress, but it moves the risk rather than removing it. Each new layer is more code that someone must read, more configuration that someone must trust. And more third-party projects that arrived last month with a very large crowd and no history. A skills file is an instruction set that your agent will follow, so a bad one is not a harmless download. A gateway sits in the path of your requests. An agent runtime holds permissions. These are exactly the places where you want slow, boring scrutiny, and the star count rewards the opposite. So the practical test is a simple one. Before adopting anything from a trending list, ask how old it is, who maintains it, what it can touch. And what happens to your data when it runs. A repository that is three weeks old and has half a million stars has told you that people are curious. It has told you nothing about whether it is safe to give it the keys. Popularity is a signal of demand. It is not an audit, and the layers now in fashion are the ones that most need one. Here is a thread running through three things I've been reading, and it is less glamorous than any of them. Alerts, lessons and star counts all share one weakness: producing the signal is cheap, and doing something sensible with it is expensive. Start with the dullest example, which is also the most useful. A vendor has published guidance on using video analytics in Brazilian factories, against two safety standards, NR-12 for machinery and NR-35 for work at height. The pitch is familiar. Point existing cameras at danger zones, let software spot someone stepping into a machine's boundary. Or working at an edge without a visible harness, and log it with a timestamp. What I find interesting is how much of the guidance is a list of things the system does not do. It cannot confirm anchorage strength. It cannot tell from video alone whether a machine is running or whether a person is authorised to be there. It does not replace guards, interlocks, energy isolation, training or permits. And the document says plainly that installing the software is not compliance. That is a refreshingly modest position for a vendor, though I would note that it is the vendor's own account. And modesty in the small print does not always survive the sales call. The part that matters most in practice is the unglamorous one. An alert is useful only when someone is responsible for acting on it. Who receives it? How is an urgent one escalated? How is the fix checked on every shift? The guidance suggests a limited pilot, to learn whether the system detects the hazards it claims to. And whether the plant has the people to respond, and then a review against actual safety outcomes before rolling it out further. Notice what that implies. The cost is not only cameras and edge hardware. It is the supervisor's time, every shift, looking at things the software has flagged. A system that raises more alerts than a plant can answer has not improved safety. It has created a record of hazards that were seen and not dealt with, which is a worse legal. And moral position than not having seen them. There is also personal data. Watching identifiable workers falls under Brazil's data protection law, so the plant needs a lawful basis, a stated purpose, limited retention and access controls. Blurring faces and avoiding facial recognition are sensible ways to reduce exposure. Workers who are being watched for their own safety will still ask what else the footage is used for. And the answer needs to be better than a policy document. Now the second piece, which looks unrelated and is not. O'Reilly has been asking what it would mean for an agent to be self-evolving, meaning one that improves without a new model. The argument I took from it is a good discipline. If an agent keeps changes from earlier rounds and builds on them, you should be able to say what it learned. What changed, and how you checked that the change helped. Then you test whether the lesson holds on a new task. The aim is an agent that needs less repeated correction. That last test is the one that gets skipped. Any system can look better on the tasks it has already seen. The honest question is whether the improvement transfers, and that is exactly the question the factory pilot is asking in a different setting. Does the system work on the next production line, not just the one it was tuned on? Same shape, different hardware. Third, the GitHub trending list for October, which I have been picking over. The analysis offered with it makes a point I think is right. The interesting work has moved above the model, into runtimes, skills and control surfaces, and below it, into cheaper inference. The model itself is treated as the one part nobody needs to own. And then comes the observation that I would underline. October produced four major skill collections, five agent runtimes and three inference projects. Against exactly one repository concerned with auditing whether any of it did what it was asked. That asymmetry is the same one as in the factory. Plenty of effort goes into detecting, generating and running. Very little goes into verifying that the response happened and the result held. One bright spot is a security-audit skill from Cloudflare that produces independently verified, machine-readable findings. Which separates it from the many skills that ask you to take the agent's word. I would still read it before running it, but the design instinct is the correct one. Here is my analysis, rather than anything the sources claim. Detection is a commodity problem. Cameras, models and agents will all get better and cheaper at noticing things and doing things. What stays scarce is the loop behind them: a named person, an escalation path, a check that the correction worked. And evidence that the lesson carries to the next case. Vendors sell the first half because it demos well. Buyers should spend their scrutiny on the second half, because that is where the outcomes come from, and. Where the liability sits when they do not. So the test is the same whether you run a plant or a fleet of coding agents. Before you add another layer that watches or acts, ask who is answerable for what it finds. How you would know the fix worked, and whether it would still work somewhere new. If nobody can answer, you have not bought safety or reliability. You have bought a very attentive witness. It has been a heavy week, with more announcements, claims and counterclaims than anyone could reasonably be expected to weigh up. That is exactly why clarity matters: most of it will not survive contact with a second reading, and it helps to know which parts will. If you would like help with that, I send out a daily artificial intelligence briefing. It is one email, no hype and no fluff, just what happened and whether it matters. You can sign up at jonathan-harris dot online. This week's sponsor is, conveniently, me. Or rather, my book: "artificial intelligence in Aviation: Transforming Safety and Sustainability". Aviation is an industry that has spent decades learning to distrust confident claims. And it has a great deal to teach the rest of us about using artificial intelligence sensibly. The book looks at where the technology helps with safety and sustainability, and where it is asked to do rather more than it should. You will find it in the eBooks section of my website, and I would be glad if you took a look. That's your lot for this week's Turing's Torch. If you want the daily brief, head to jonathan-harris dot online. Same time next week — try not to believe the press releases.