The drug candidate waiting for validation is the image I’d keep from this report. McKinsey describes a similar backlog between generated code and secure deployment, but my reading is that cheaper generation makes selection more valuable - the demonstration ends just where the expensive judgement begins. For a software team, I’d test that by capping work entering review while introducing faster generation tools. Measure time to accepted, working changes and the effort spent rejecting or repairing suggestions. If the team produces more code without delivering more useful software, the experiment has bought a larger queue. Calling that productivity would require a generous definition. But that suggests a less glamorous use for the next AI investment: helping teams reject unsuitable work before it consumes scarce review time. Better filtering could deserve priority over another increase in generation speed, provided it doesn’t quietly discard the valuable, unusual ideas too. I’d judge the next coding pilot by trustworthy changes reaching users and the rework they create. The output counter can look after itself.

This is soooo cool! But seven football pitches is a striking amount of space to set aside for a battery. This Energy Dome has traded the geological lottery of underground compressed-air storage for an enormous object you can build above ground. That freedom could make a difference, especially in regions without the right caverns. It also means planning permission, suitable land and nearby grid infrastructure deserve as much attention as the turbine. The 70% round-trip efficiency looks distinctly ordinary beside lithium-ion’s 90%. I don't think that comparison settles anything. A storage plant intended to cover a long evening of low wind competes on delivered electricity, installation cost and how many hours it can run. The company's claimed 10% to 15% cost advantage at eight hours is interesting, but I'd want that compared on a full-project basis, including land and civil works. A cheap tank doesn't make a cheap site. The Sardinian plant is tangible evidence that the machinery can be assembled at commercial scale. Thirty gigawatt-hours across five continents is a different sort of claim. Energy Dome's future probably depends on repeating a large industrial project in places with different permitting systems, contractors and grid constraints. The turbine is easy to picture. The thirtieth planning application is where my imagination stalls.

The full stop after "P.Byrnes." is the most telling character in Andrew Deck's investigation. Pat Byrnes has signed his cartoons that way since he was a teenager. In more than a dozen generated images bearing his name, the dot survived too. ChatGPT can miss the joke and still reproduce the part that tells a reader who made it. That's an impressively precise way to be wrong. When someone requests a New Yorker-style cartoon, the requested style is recognisable: a spare drawing, an awkward social scene, a dry line beneath it. A real artist's signature makes a different promise. It says this particular person chose to publish this particular thought. Emily Flake's analogy to being misquoted captures the harm better than most debates about style imitation. People may forget whether the joke was funny. They may remember who supposedly told it. The design implication seems straightforward. Generated images shouldn't acquire someone else's signature by accident, and a warning about stylistic similarity misses the precise failure here. After Deck's notification, ChatGPT began showing similarity warnings, yet signed outputs persisted. A creator shouldn't have to audit social media for punchlines they've never written. The small mark in the corner is doing far more reputational work than the rest of the picture.

Three hundred and ten times is a number built for a headline... But that's the reported increase in GLP-1 prescriptions for US children aged eight to 11 with obesity between 2019 and 2026. Jessica also adds the less photogenic detail: these prescriptions remain rare; without an absolute starting figure, the multiplier tells us how rapidly prescribing is changing while revealing little about how many children are being treated. The difficult part is the mismatch in timescales. Adult side-effect studies can record nausea or nail changes over relatively short periods. Understanding what a drug might mean for a growing child takes years. The WHO's new guidance declines to recommend medication below ten because evidence about growth, development and mental health remains thin. So, that caution doesn't resolve every clinical decision when a child already faces health problems. Yet, I'd favour following treated children through developmental milestones, documenting why treatment began, what prompted stopping it, and outcomes after discontinuation. Hair loss and nail changes make arresting headlines, but the unknowns that take years to emerge are harder to photograph. GLP-1 press releases arrive almost daily, and childhood isn't going to run on the industry's publishing schedule....

We're back to finding weird projects across the internet, again :) Flour-covered hands are a better test of ambient computing than another floating dashboard! In Guillaume kitchen demonstration, the interface follows the job: a timer and recipe on the worktop, with no need to retrieve a phone while handling food. It sounds trivial until you consider how much interface design assumes a clean finger, a free hand and permission to interrupt whatever you're doing. The shared travel map is the more consequential demonstration - but rather cool. Two people can point at the same place, see the same information and follow each other's gestures. Nobody needs to narrate what's happening on their private display. I'd like to see this approach explored on repair benches, in classrooms and around workshop tables, where physical surfaces already organise collaborative work. The computer becomes another participant in an existing activity. Yet, there's an interesting social trade-off here; a device that sees an entire room and projects information onto its surfaces needs to distinguish between what people want to share and what they'd prefer to keep to themselves. He proposes hardware safeguards, although these remain design intentions rather than demonstrated protections. Getting the interaction right will require understanding the room's occupants as carefully as its geometry.