Overview
Today’s feed had two clear threads: AI moving from demos to day-to-day tools (coding agents on your phone, hardware-aware transformer training, and even dating apps handing choices to a matchmaker), and big, messy geopolitics where chips, oil, and boardroom photos end up part of the same story. Add humanoid robots hitting endurance milestones, plus a reminder that the algorithm still runs the show on social feeds.
The big picture
We’re watching automation spread sideways. Not just smarter models, but better packaging: remote control for agents, courses that teach the bits that break in production, consumer apps trying to swap endless scrolling for “just decide for me”, and robotics proving it can keep going past the short demo window. At the same time, the biggest constraints are looking less technical and more political, with AI hardware caught up in summit diplomacy and sanctions talk.
Codex comes to your phone, while it runs at home
OpenAI is previewing Codex inside the ChatGPT mobile app. The idea is simple: kick off work, check progress, review outputs (think diffs and logs), then approve the next step, all from iOS or Android, while the agent keeps running on your laptop or dev box.
It’s a small product change with big behavioural impact. If this works smoothly, it turns long-running coding tasks into something you can supervise in spare minutes, instead of babysitting a machine that cannot be allowed to sleep.
Learning transformers with the parts that actually bite in production
Andrew Ng announced a new course, “Transformers in Practice”, built with AMD and taught by Sharon Zhou. It’s pitched at the practical layer: how tokens get generated, what attention is doing, and how to spot and fix things like slow inference.
The useful signal here is the focus on deployment judgement, not theory for its own sake. People building real systems keep running into the same questions - why outputs go off the rails, where latency appears, and which optimisation trade-offs are worth it.
Humanoid robots edge towards the boring part, which is the point
Brett Adcock says Figure’s humanoid robots are nearing 24 hours of continuous package sorting, with “Bob, Frank, and Gary” past 28,000 packages and still running. That endurance claim matters more than a flashy five-minute clip, because factories do not care about demos, they care about staying up.
If the reliability holds, the next questions come quickly: error rates, battery strategy, maintenance, and how much autonomy is real versus carefully staged conditions.
AI and geopolitics: Jensen Huang’s summit comments raise the temperature
Watcher.Guru quoted Nvidia CEO Jensen Huang calling the Trump-Xi summit “one of the most important summits in human history”. The dramatic line lands because Nvidia sits right at the pressure point: national policy on chips versus commercial demand in China.
The subtext is familiar now. AI leadership arguments often end up expressed as export rules, but the winners and losers are not evenly shared across “America” as a category.
That Air Force One photo op says plenty without saying anything
Sawyer Merritt posted photos of Elon Musk and Jensen Huang chatting on Air Force One. It’s the kind of image that becomes a Rorschach test: influence, access, trade priorities, and who gets to be in the room when policy gets written.
Even if nothing concrete comes from a single flight, it’s a neat snapshot of how AI hardware, industry leaders, and state diplomacy have started travelling together.
Sanctions watch: Trump hints at action on Chinese buyers of Iranian oil
The Kobeissi Letter flagged Trump saying he will soon decide what to do about Chinese companies buying Iranian oil. Markets tend to treat this as a headline risk until it is not, then price moves arrive fast.
It also links back to the broader US-China bargaining table: energy enforcement can become another lever in negotiations that already include tech and chips.
Bumble wants to bin swiping and hand decisions to an AI matchmaker
Polymarket says Bumble plans to remove left-right swipes and replace them with an AI matchmaker called “Bee”. If true, it’s a bold admission that the swipe mechanic has worn people down, even if it once defined the category.
The test will be whether users trust a machine to pick candidates, and whether the app can prove it is improving outcomes rather than just changing the interface.
The nastiest growth hack is still… the App Store UI
Nikita Bier shared a favourite trick: naming the company “Tap Get Inc” so the developer label beside the download button reads “Tap Get”. It’s mischievous, simple, and hard to unsee once you know it exists.
It’s also a reminder that distribution does not always come from grand strategy. Sometimes it comes from noticing a single line of interface text and treating it like prime real estate.
If the next outbreak hits, the response playbook could look nothing like last time
Bryan Johnson sketched a hypothetical: a mutated hantavirus triggering an AI-plus-biotech sprint - rapid sequencing, protein target mapping, drug screening, and parallel vaccine design. It’s part provocation, part optimism about how fast the pipeline could run with modern tools.
The interesting part is less the exact timeline and more the assumption that capability will be distributed. Not just governments and big pharma, but a wider mesh of labs, platforms, and manufacturing.
The timeline is still a slot machine, and it knows what you clicked
Trung Phan posted a meme about liking a single Ed Harris post and then getting flooded with Ed Harris content. It’s a joke that lands because everyone has lived it: one tap, then your feed turns into a hall of mirrors.
Underneath the humour is the same old lesson. Recommendations are not “what’s popular”, they’re what the system thinks will keep you leaning forward.
