3 billion iphones📱 hitchhiker's guide to the AI bubble🫧 personal superintelligence👓🪄 genie 3🧞 claude 4.1💬 ad‑lexa🛒 de novo🆕 wide research🕸️ darwin monkey🐒 study mode🎓 money2🏦 #2025.31
Store data on a bird
Welcome to this week's Memia scan across AI, emerging tech and the exponentially accelerating future. As always, thanks for being here!
This week’s issue comes to you from Cambridge, UK - I’m back in the UK briefly before heading off down through Central Europe next week. Reach out if you’re around Wroclaw, Prague or Vienna during August to catchup over coffee…
ℹ️PSA: Memia sends *very long emails*, best viewed online or in the Substack app.
🗞️Weekly roundup
The most clicked link in last week’s newsletter was Google Deepmind CEO Demis Hassabis in conversation with Lex Fridman: Future of AI, Simulating Reality, Physics and Video Games.
ICYMI
My recent post on macro AI uncertainties:
Memia RSS feed
I keep chipping away at the Sensorium feature backlog … now you can integrate Memia’s real-time AI and tech news scanning feed into your RSS reader if you use one. All the links which I’m sharing throughout the week are now available here: https://www.s7m.net/links/rss
Follow-ons from last week
A few developments on stories covered in last week’s issue:
China’s call for Global AI Governance
Commentary arising from the recent World AI Conference (WAIC) in Shanghai:
WIRED: China leads global AI safety push as US steps back.
“With the US out of the picture, a coalition of major AI safety players, co-led by China, Singapore, the UK, and the EU, will now drive efforts to construct guardrails around frontier AI model development…
…A competitive landscape of multiple powerful open-source AIs is in the best interest of AI safety and humanity's future … because different AIs naturally embody different values and will keep each other in check.”
TechXPlore: 'Marathon at F1 speed': China bids to lap US in AI leadership.
🇬🇧🔒The UK’s Online “Safety” Act
Politico: The UK’s new tech law triggers upheaval.
“By using a VPN to simulate UK web browsing, DFD was able to confirm reports that content relating to Gaza on X and cigars on Reddit was more restricted in the UK than in the US“
The Register: Banning VPNs to protect kids? Good luck with that.
“The government could pull various technical levers, such as banning the sale of VPN kit, but … it would be like banning people from smoking in their own homes.
"You might not like it, but good luck enforcing it," said Graeme Stewart, head of public sector at Check Point Software. "The logistics are near-impossible. You could, in theory, ban the sale of VPN equipment, or instruct ISPs not to accept VPN traffic. But even then, people will find workarounds. All you'd achieve is pushing VPN use underground, creating a black market for VPN concentrators.
"The only way to do it is badly. You'd effectively be forcing ISPs to block legitimate encrypted traffic and, in doing so, you'd be regulating an entire industry out of existence. Worse still, you'd be legislating against cybersecurity and privacy…
…Some countries that ban the use of VPNs include Russia, the United Arab Emirates, Iran, Saudi Arabia, Turkmenistan, Myanmar, Belarus, and China. That's not even an exhaustive list, but it shows the questionable company the UK would keep should it choose to ban VPNs.”
Top-of-the-class response here:
Meanwhile the age-gating debate is hotting up elsewhere around the world right now:
TechSpot: Meta, YouTube, TikTok fight back as Australia's social media restrictions for teens draw closer
TechCrunch:
YouTube rolls out age-estimation tech to identify US teens and apply additional protections YouTube is rolling out machine learning–based age estimation in the US to flag likely teen users regardless of self‑reported birthdates,
Google is experimenting with machine learning-powered age-estimation tech in the US
Narrative: Minority Report vibes all round. A fundamental societal policy choice between privacy and secure communications vs. state surveillance.
As China’s experience has shown, increasingly sophisticated VPNs constantly evolve to find a way to evade detection. Fundamentally, information is unconstrained by geographical borders and will find a way to flow around attempted barriers. Better for democratic governments to acknowledge that up-front and explore other approaches to protecting citizens (from themselves).
Delta Airlines’ “personalised pricing”
Techspot: ✈️ Delta denied AI-driven personalised “surveillance pricing” amid US Senate scrutiny. But lawmakers remain sceptical, warning that “aggregate” signals can stealth-personalise, and floating federal rules to ban AI pricing based on personal data. (LOL - the US is so mixed-up…)
Tea app security breach
TechXPlore: Tea app takes messaging system offline after second security issue reported Tea — the “women-first” doxxing app that surged to the top of the US App Store — confirmed that some user DMs were also accessed in its recent breach and as a result it has taken its messaging system offline. This came days after acknowledging exposure of roughly 72,000 images (including 13,000 verification selfies/IDs). Oof.
📈 Thinking about AI
Two pieces on AI which caught my eye this week:
🆕De novo
An FT profile piece on US startup OffDeal contains a nugget of AI Zeitgeist gold. The firm is building a niche sell-side M&A “investment bank” using AI, targeting smaller US SMB clients than traditional firms and swapping out 4–5 person deal teams for agentic AI workflows that source buyers, drafts teasers/CIMs, run NDAs/data rooms and manage bidder comms end-to-end.
Co-founder Ori Eldarov is quoted:
“I came up with a crazy idea, like, “[w]hat if I just start my own investment bank? Like, what if I just built a de novo bank from scratch ... my own software, my own org structure, my own data, my own bankers ... a fully vertically integrated kind of approach”
Signal: replace “de novo bank” with “de novo [insurer | consulting firm | accountant | fund manager | telco | advertising agency …"]. Whatever the niche, I guarantee there are entrepreneurs out there executing on that right now using AI.
🫧Hitchhiker's Guide to the AI Bubble
A superb post on LinkedIn: A Hitchhiker's Guide to the AI Bubble by UK tech strategist Django Beatty.
One of the most uncluttered and concise deconstructions of the current AI “Bubble” moment. Lots of gems in the post, highly recommend reading the whole thing, but here are the key passages:
"That 16:1 AI investment ratio only makes sense if the winner takes everything.
Which is exactly what everyone believes…
…This is your bubble: Not the technology - [but] the shared delusion that someone's about to achieve irreversible computational dominance.
...The AGI bubble will pop. Not because the technology fails, but because the fantasy can't survive contact with reality.
The trigger could be anything. A major AI company admitting they're nowhere near AGI. A government realising they've been stockpiling GPUs for a race that doesn't exist. Or simply investors noticing that $560 billion for $35 billion in revenue isn't a business model - it's a cargo cult.When it happens, the narrative collapse will be spectacular. All those breathless headlines about consciousness and superintelligence will age like dot-com era predictions about the "new economy" where profits didn't matter. The Stargate project will become this generation's Webvan - ambitious, well-funded, and built on false premises.
But here's what the doomsayers miss: the infrastructure remains. After the dot-com crash, we still had fibre optic cables, data centres, and trained engineers. The speculation died. The internet didn't.
Same pattern here. When the AGI fantasy evaporates, we'll still have:
Models that can read, write, and analyze
APIs that cost pennies to call
A generation of developers who know how to build with them
Actual products solving actual problems
The companies that survive won't be the ones promising AGI. They'll be the ones who understood early that ML is just really useful when available as infrastructure. Like the difference between Pets.com and Amazon - one promised to change the world, the other was building warehouses.”
It’s compelling reasoning. I’m increasingly persuaded that the Dotcom bubble analogy is the right one… so the question then becomes: how long until the AI bubble goes “pop” … surely any day now…?
(And: what will you build on the other side….?)
🏭AI industry news
Speaking of AI bubbles…it’s been a huge week of giant US tech company earnings announcements, huge AI capex splurges… too many to cover in depth so summarising the main headlines / narratives for each company into a few links:
Nvidia
🛡️China accuses Nvidia of backdoor in H20 chip China's cyber regulator has summoned Nvidia over alleged "serious security issues" with its H20 chips, claiming experts revealed the hardware contains location tracking and remote shutdown capabilities. This “backdoor diplomacy” curveball comes just weeks after Washington lifted its ban on H20 sales to China and Nvidia CEO Jensen Huang made a charm offensive trip to Beijing to rebuild the company's business there.
The claims are pretty vague and Nvidia flatly denies any remote access features, — however the accusations may preempt US lawmakers who are weighing a mandate to embed tracking in export‑controlled silicon.
The timing — just after the H20 green light — gives this a feel of strategic theatre to maintain momentum for China's domestic chip transition, giving a boost to domestic contenders like Huawei’s Ascend 910C.
As the first responder to my post on X put it:
Meta
🧠👓🪄Personal superintelligence for all (so long as you watch ads)
Hours before Meta’s earnings announcement, Mark Zuckerberg dropped a manifesto pitching “personal superintelligence” for everyone that maximises leisure time: more AI-personalised feeds, AI-generated Reels, and parasocial chats with AI personas, all monetised with precision ads.
“As profound as the abundance produced by AI may one day be, an even more meaningful impact on our lives will likely come from everyone having a personal superintelligence that helps you achieve your goals, create what you want to see in the world, experience any adventure, be a better friend to those you care about, and grow to become the person you aspire to be.
Meta's vision is to bring personal superintelligence to everyone. We believe in putting this power in people's hands to direct it towards what they value in their own lives.”
(🤢)
Zuckerberg also leaned in strongly on AI-native smart glasses ase the primary way we will interact with superintelligent assistants—letting them “see” and “hear” your world and going so far as warning that those without them will be at a “cognitive disadvantage”. (Reminder he needs to justify Reality Labs still burning through cash — US$4.53B loss in Q2; nearly US$100B since Meta acquired Oculus in 2014). Nevertheless, its a valid hypothesis.
So “Personal Superintelligence” was the vision pitched on the earnings call, backed by the new Superintelligence Labs team formed out of the US$14.3B acquihire of Scale AI’s Alexandr Wang and eye-watering comp offers to other researchers.
More mundanely, Meta credited upgrades to its AI ranking engines boosted engagement in Q2, nudging time spent up 5% on Facebook and 6% on Instagram, while video minutes jumped 20% year-over-year. The family of apps hit 3.4 billion daily users and US$47.1b in revenue, with Threads’ stickiness also rising thanks to LLM integration.
Meta also announced it is cranking AI infrastructure capex up again to US$66–72B in 2025—with another step-up planned in 2026— anchored by its Prometheus (1 GW, Ohio) and Hyperion (up to 5 GW, Louisiana) data centres, while exploring external financing to co-develop even larger “mega” data centres.
Investors shrugged and sent the stock up ~10% after-hours. Go figure.
The earnings announcements dissected:TechCrunch: Meta to spend up to $72B on AI infrastructure in 2025 as compute arms race escalates
TechCrunch: Zuckerberg says people without AI glasses will be at a disadvantage in the future
The Verge: Mark Zuckerberg promises you can trust him with superintelligent AI …And he’s still really bullish on smart glasses.
The Verge Command Line: Decoding Zuckerberg’s ‘personal superintelligence’ plan: fill your free time with more AI
TechCrunch again: Zuckerberg signals Meta won’t open source all of its ‘superintelligence’ AI models
Oh, and Zuck has been targeting Mira Murati’s 50-person Thinking Machines Lab team with billion-dollar AI offers … and apparently getting turned down.
Amazon
It’s a topsy-turvy world: Amazon’s quarterly sales results also beat expectations, but shares dropped 7% as its AWS cloud computing arm’s growth rode the gen‑AI wave (sales +17.5% to US$30.9b) but apparently lagged behind rivals. Trump’s tariffs weigh heavily on Amazon’s ecommerce outlook…
🗣️🛒 Ad‑lexa Amazon wants Alexa+ to slip “tasteful” ads into your multi‑turn chats, betting conversational discovery can help fund its US$31.4B Q2 AI capex binge. 22%. The assistant’s new tiers (Prime‑bundled, US$20/month standalone, likely ad‑free to come) signal a freemium model where attention, intent and voice data become the real monetisation layer.
Narrative: Who would willingly buy a talking AI salesperson for their own home? Dark.
Microsoft
Microsoft hits US$4 trillion Microsoft’s FY2025 Q4 was a cloud flex: the company joined Nvidia in the US$4 trillion valuation lounge after disclosing Azure surged 39% YoY and pulled in US$75 billion in FY2025 and posting stronger-than-expected earnings. The stock jumped ~4% as investors priced in a world where Azure quietly powers… everything?
Also: CFO Amy Hood signalled an eye-watering US$30 billion on AI infra next quarter alone.
📜Microsoft seeks AGI-proof OpenAI deal, bigger equity stake
Microsoft is reportedly negotiating a new deal to guarantee access to OpenAI’s models *even if* the startup declares “AGI” while paving the way for OpenAI’s shift from its nonprofit/capped‑profit hybrid to a fully commercial structure. The talks include Microsoft taking a low‑ to mid‑30% equity stake, extending rights past the current 2030/AGI cutoff, and formalising safety commitments as capabilities scale. The wild cards: antitrust scrutiny and Elon Musk’s lawsuit (plus the small matter that no one agrees on what “AGI” means).
OpenAI
Speaking of OpenAI…
OpenAI raises $8.3bn in massive funding round OpenAI just closed a US$8.3bn funding round as part of a larger $40bn raise, with the company now valued at US$300bn and boasting US$12bn in annual recurring revenue. Dragoneer Investment Group wrote a jaw-dropping US$2.8bn cheque—one of the largest single startup investments ever—while the round was five times oversubscribed with participation from the usual Silicon Valley suspects.
Apparently the company has now hit US$12B annualised revenue and a claimed 700M weekly ChatGPT users.⚡OpenAI brings 290MW renewable-powered data center to Norway OpenAI is extending its “Stargate” AI data centre footprint to Norway with a 290‑megawatt, renewables‑powered installation
OpenAI halts experiment exposing shared ChatGPT chats to search OpenAI pulled the “discoverable by search” experiment after shared ChatGPT conversations (chatgpt.com/share) starting turning up fully discoverable in Google Search. Sharing was opt-in with a discoverability toggle, but many users didn’t get that “public link” plus web crawlers equals *full visibility*. (Imagine: a screenshot of Google results showing chatgpt.com/share links with snippets like “Rewrite my resume for [Company]…”.)
Perplexity
At least Google respects “do not index” instructions…
Cloudflare says Perplexity evaded robots.txt to scrape sites Cloudflare says Perplexity has been scraping sites that explicitly opted out via robots.txt, allegedly masking its identity by rotating user agents at scale across millions of requests. Perplexity denies the claims, calling Cloudflare’s post a sales pitch for its marketplace to charge AI scrapers.
Let the games begin…
Anthropic
💸Anthropic nears $170B valuation with up to $5B raise Anthropic is reportedly closing a US$3–5B round led by Iconiq Capital that would value the company at ~US$170B, nearly tripling its March mark of US$61.5B. Talks have included Qatar Investment Authority and Singapore’s GIC, signalling that sovereign wealth funds are sprinting into frontier AI (at the top of the bubble, though…?). Cue the awkward memo: CEO Dario Amodei admits discomfort with “dictatorial” money while conceding the physics of capital—because who else can bankroll compute that now looks more like national infrastructure than a startup budget?
Notably, Anthropic is now open to Middle East investment despite prior reluctance—an ethical pivot from CEO Dario Amodei.🏢Anthropic overtakes OpenAI in enterprise AI market share
Anthropic now leads enterprise LLM usage at 32% versus OpenAI’s 25%, per Menlo Ventures, a sharp reversal from 2023 when OpenAI held 50% and Anthropic 12%. In coding workloads the gap is wider: Claude commands 42% vs ChatGPT’s 21%, buoyed by Claude 3.5 Sonnet (Jun 2024) and 3.7 Sonnet (Feb 2025), while Google inches up and Meta remains the (American) open source leader.1
🛑🔌 API Moats: Anthropic cuts OpenAI off from Claude
Anthropic revoked OpenAI’s Claude API access, alleging terms-of-service violations after finding OpenAI hooked Claude into internal tools and coding workflows ahead of its upcoming GPT-5 release, which Anthropic says breaches its no-building-competing-services ToS. A carve-out remains for benchmarking and safety evaluations, while OpenAI calls its usage “industry standard” and notes its own API stays open to Anthropic.
Mistral
🇫🇷💶Mistral eyes US$10bn valuation to fuel AI expansion
Paris-based Mistral is raising about US$1bn at a US$10bn valuation to scale its “more open” LLMs and Le Chat chatbot, while accelerating an €8.5bn data centre build outside Paris with MGX, Nvidia and Bpifrance. Revenues are set to top US$100m/year, buoyed by a handful of ≥US$100m multi‑year deals across finance, logistics and France’s defence ministry. Europe’s champion still trails OpenAI (US$300bn) and Anthropic (targeting US$170bn valuation).
Google
Alphabet beats earnings expectations, raises spending forecast Just like Microsoft and Amazon, Google’s parent company announced second-quarter results last week that showed huge increases in cloud revenues, with overall revenue up 14% year over year, far higher than 10.9% analyst predictions.
Google Cloud revenue of $13.62 billion was up 32% YoY.
The company also increased its capex forecast for 2025 to US$85 billion, up US$10 billion from February, due to “strong and growing demand for our Cloud products and services.”
🧭Google backs EU AI code as Meta holds out Google will sign the EU’s code of practice for general-purpose AI—aligned with the bloc’s incoming AI Act—joining OpenAI and Mistral as Brussels doubles down on safety, copyright and training-data transparency. Meta is still refusing, calling the code overreach that creates legal uncertainty and could throttle frontier models in Europe, while Microsoft hints it’s likely to sign.
⚖️Court upholds Epic win: Google Play deemed illegal monopoly
In the US, the Ninth Circuit upheld a jury finding that Google’s Play Store is an illegal monopoly, forcing Google to allow rival app stores inside Play, permit alternative billing, and make its full app catalogue available to competing marketplaces (with developer opt-outs). It’s running out of time to comply.
📱Google mocks Apple’s delayed Siri in Pixel 10 ad this is uncharacteristic shade from Google:
Apple
Ah…. Apple.
📱📈🌍 Three billion iPhones
Apple passed 3 billion iPhones sold since 2007, matching its first-billion pace by adding another 2 billion in the last nine years, while Q3 iPhone revenue jumped 13% year-over-year to US$44.6B out of total revenue this quarter of US$94B.
Despite the milestone, Apple trails Nvidia and Microsoft in market cap, and analysts suggest the sales spike may have been pulled forward by tariff fears.🧾Apple says Trump tariffs will add US$1.1 billion this quarter Apple says Trump-era tariffs will tack on another US$1.1 billion to its September-quarter costs after burning through about US$800 million in June, as duties hit devices made across China, India, and Vietnam. Tim Cook flagged that most iPhones sold in the US now come from India, while Macs, iPads, and Watches skew Vietnam.
🧩Tim Cook says Apple open to AI acquisitions… The Apple CEO said that the company is open to AI acquisitions while reallocating more staff to ship intelligence across devices and platforms. With Siri’s big revamp delayed for *reliability*, Apple is reportedly courting OpenAI and Anthropic and even weighed buying Perplexity.
Tim Cook on AI: “Apple must do this” In a rare all-hands, Tim Cook called AI “as big or bigger” than the internet and smartphones, promising Apple will invest heavily while rebooting its Siri overhaul onto a single, LLM‑centred architecture. Software chief Craig Federighi said a hybrid legacy/LLM approach failed Apple’s quality bar, so the team is migrating “everything” to the new stack. Plus, talent drifting to Meta…
🔎Apple building answer engine? Apple is rumoured to be assembling an AI “answer engine” under its AKI (Answers, Knowledge and Information) team to crawl the web and deliver direct responses to challenge Google Search.
Looks like the Perplexity deal didn’t progress after all…
Palantir
More gravity-defying financials…
Data intelligence group Palantir announced quarterly revenues of US$1Bn. CEO Alex Karp waxing modest as always:
“We continue to see the astonishing impact of AI leverage…the growth rate of our business has accelerated radically, after years of investment on our part and derision by some”.
With forecast annual revenues for 2025 now slightly at US$4.1Bn, somehow the market cap is sitting at US$392Bn - nearly 100X revenue multiple!!!
Tesla
⚖️🚗🛑 Tesla’s US$240M Autopilot Reckoning
A Miami jury has decided that Elon Musk's car company Tesla was partly responsible for a deadly crash in Florida involving its Autopilot driver assist technology and must pay the victims more than US$240 million in damages. The 2019 Florida crash involved a Model S on Autopilot blowing through a flashing-red/stop at about 97–113 km/h, killing one person and severely injuring another. Plaintiffs argued that Tesla’s design lacked geofencing and its marketing overstated safety, while Tesla blamed the driver.
Narrative: This case is a bellwether for how regulators, courts, and insurers will rewrite the autonomy stack’s real-world constraint model—less about sensors and compute, more about liability, language, and clarifying where responsibility sits between manufacturer, machine and human operator.
🏭Samsung’s $16.5bn Tesla chip deal tests foundry revival
Samsung landed an eight-year, US$16.5bn deal to build Tesla’s next‑gen AI6 chips at its delayed Taylor, Texas fab—a lifeline for a foundry arm stuck at 7.7% market share while TSMC sits at 67.6%.
Waymo
🤖🚕🛠️ Waymo x Avis in Dallas
Waymo will launch its autonomous ride-hailing service in Dallas next year, tapping a multi‑year partnership with Avis to run depots and maintenance while Waymo focuses on the Driver stack riding atop Jaguar I‑PACE EVs.
Walmart
Walmart consolidates dozens of AI chatbots into four 'super agents'
Walmart is consolidating dozens of fragmented AI tools into four comprehensive "super agents" - customer-facing Sparky, employee-focused Associate, supplier-serving Marty, plus a developer platform - aiming to drive online sales to 50% of its US$648 billion revenue within five years.
Narrative: What do we take from everything above? (1) AI is driving hyperscale cloud revenues up ~30% YoY and (2) tech giants and their investors still think there’s plenty more AI gravy to come. But viewed in a frame of the AI Bubble popping, perhaps Nvidia and the 3 hyperscale cloud giants will be the most resilient as they are sitting on long term infrastructure capacity…. but the pumped-up valuations of pure-play AI companies feel *very* frothy right now.
⚡💧Hungry and thirsty data centres
Four stories on data centre (un)sustainability:
⚡Wyoming AI data centre could outpower entire state’s homes
The US state of Wyoming is weighing a Tallgrass–Crusoe AI compute campus near Cheyenne that would draw 1.8 GW in phase one (about 15.8 TWh/year) and scale to 10 GW (87.6 TWh/year) — more than all Wyoming households use and ultimately more than the state produces. The developers propose dedicated natural gas plus renewables, while a mooted nuclear plant sits over the horizon, and the anchor tenant remains unnamed amid OpenAI “Stargate” speculation.
🏢🇪🇺Datacentre giants challenge EU efficiency rules
Europe's largest datacentre operators, including AWS, Microsoft, and Google, are pushing back against EU proposals for minimum performance standards, arguing the regulations are premature and based on flawed data collection. The Climate Neutral Data Centre Pact claims the rules could "cement flawed metrics into long-term policy" and undermine Europe's AI competitiveness agenda.
Yeah right: with individual AI datacentres now planned to consume more electricity than entire US states, one wonders if "let us self-regulate a bit longer" is really the optimal strategy here? Rigorous transparency needed.
🏜️Gulf’s AI boom collides with a looming water crisis
The UAE and Saudi Arabia are building gigawatt-scale AI campuses to lure OpenAI, Google, Microsoft et al, but data centres in the region could drink 426 billion litres a year by 2030, with desalination-fuelled grids already creaking. Mitigations from treated wastewater to closed-loop and air-cooled designs exist, yet deployment lags and transparency on water metrics is sparse, creating a credibility gap with glossy sustainability pledges.
💧Texas data centres drain 50 billion gallons amid drought
Likewise, Texas data centers consumed 49 billion gallons of water in 2025—enough for millions of households—with usage projected to explode to 399 billion gallons by 2030 as AI infrastructure expands. Meanwhile, San Antonio residents face once-weekly lawn watering restrictions during ongoing drought conditions, while local Microsoft and Army Corps facilities alone drained 463 million gallons over two years.
Narrative: Welcome to the AI-ocene Epoch, where AI terraforms to the planet to its needs, not humans’.
🆕 AI releases
One of those weeks when there’s a whole slew of mould-breaking AI models and tools…
⚡Bolt Zeus GPU
Chip startup Bolt Graphics is hyping “Zeus” — its GPU family that it claims outperforms Nvidia’s RTX 5090 in path tracing while sipping far less power, packing gargantuan caches and socketable VRAM that scales to terabytes. The fine print: it’s all pre‑silicon, simulated estimates — the real hardware won’t land before 2027. (And it won’t melt, honest…).
💼GLM-4.5
Chinese startup Z.ai launched GLM-4.5 and GLM-4.5-Air, open-weights models. GLM-4.5 is a 355B parameter MoE model that claims to unify reasoning, coding, and agentic capabilities into a single system, ranking 3rd overall in benchmarks against OpenAI, Anthropic, and other major players.
The Apache 2.0-licensed models undercut Western API pricing significantly and run efficiently on just eight Nvidia H20 GPUs—half the hardware requirements of comparable models.
Z’s party trick is automatically generating complete PowerPoint presentations from simple prompts:
🪄 gpt-oss
Just as I was writing the newsletter, OpenAI dropped its first highly-anticipated open‑weight reasoning models: gpt‑oss‑120b and gpt‑oss‑20b, on Hugging Face under Apache 2.0.
120b runs on a single Nvidia GPU and 20b on a 16GB laptop
They’re text‑only locally but “agentic”, MoE‑based, RL‑tuned, and can route complex or multimodal queries to OpenAI’s closed cloud models, effectively making the “open” tier a smart front‑end to the proprietary stack.
They benchmark ahead of leading open peers while trailing OpenAI’s closed o‑series
Full OpenAI model card here.
Early analysis from Nathan Lambert here:
🎓 ChatGPT study mode
OpenAI rolled out a Socratic “study mode” in ChatGPT, nudging students with open-ended prompts to reason their way to answers instead of copy-pasting them. Built with input from educators and cognitive scientists, it’s available to most logged‑in users (including the free tier) and personalises by grade level and goals.
🌶️Spicy mode xAI launches Grok Imagine with NSFW ‘spicy mode’ on X
(Just because you *can*, doesn’t mean you *should*. Who is the target audience for Grok…? The product management decisions keep bemusing me… who wants this?)
🧮Gemini 2.5 Deep Think Google DeepMind shipped Gemini 2.5 Deep Think, its first public multi‑agent reasoning model, to Ultra subscribers (US$250/month), claiming state‑of‑the‑art scores on Humanity’s Last Exam and LiveCodeBench while orchestrating parallel “agent swarms” with built‑in tools like code execution and Search. A sister variant that helped clinch a gold at the International Math Olympiad will go to select academics, though it takes hours to reason—trading latency and compute for deeper chains of thought.
Google 🌍AlphaEarth Foundations
Google DeepMind’s AlphaEarth Foundations is new research which helps map the planet in unprecedented detail. The model compresses petabytes of multi-sensor Earth observation satellite data into “embedding fields” — compact, 10 m-resolution summaries that cut errors ~23.9% and storage needs 16×, enabling on-demand, cloud-robust maps through Google Earth Engine. The STP (Space Time Precision) architecture models continuous time to interpolate through cloud cover and data gaps, powering real deployments from Amazon deforestation tracking to a global ecosystems atlas.
If we can now query the planet like a vector database, how fast do environmental audits, supply chain and urban planning flip from periodic surveys to streaming observability?
🎬NotebookLM AI-narrated video slideshows
Google’s NotebookLM now spins your saved documents into AI‑narrated slideshows called Video Overviews, auto‑generating visuals and pulling in quotes, numbers, and diagrams to make data, processes, and abstract ideas more digestible. It’s launching in English first with more languages promised:
🧞BREAKING: Genie 3 Just as I was finishing the newsletter, yet another frontier model from DeepMind: Genie 3 enables “Prompt-to-world” generation, using generative AI to create a 3D “world” which unfolds in real time as you explore it.
“a general purpose world model that can generate an unprecedented diversity of interactive environments.
Given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p.”
Wow…. we’ve seen Meta and other labs working on far more basic similar efforts for a few years… but those demos are remarkable. Let that one sink in: a whole new creative palette… and will plain old reality ever be interesting ever again?
🐆Anthropic readies Claude Opus 4.1
Anthropic’s configs just outed “claude-leopard-v2-02-prod” and a nod to Claude Opus 4.1, pitched internally as the “latest model for more problem-solving power” — read: a reasoning/planning bump rather than just a vibes tune-up. Neptune v4, their safety system, is in red-teaming now, which usually puts public rollout a week or two after validation.
BREAKING: Claude Opus 4.1 released … upgrades to Sonnet and Haiku coming soon…
🤝Magentic-UI
Microsoft Research has released Magentic-UI, an open-source web interface designed to enable effective human-agent collaboration through six key interaction mechanisms: co-planning, co-tasking, action guards, answer verification, memory, and multi-tasking.
The system achieved a 71% improvement in task completion rates when humans provided lightweight feedback compared to fully autonomous operation, demonstrating that strategic human involvement can unlock productivity from imperfect AI systems.
What's particularly intriguing is their finding that even minimal human input—occurring in only 10% of tasks—can dramatically improve outcomes, suggesting we might be approaching collaboration rather than replacement as the optimal human-AI paradigm.🕸️Wide Research
Manus Singapore AI agent startup Manus launched “Wide Research” an experimental mode that spins up 100+ fully‑fledged AI subagents in parallel—each running on its own VM
“Whether you're exploring the Fortune 500, comparing top MBA programs, or diving into GenAI tools, Wide Research makes deep, high-volume research effortless.”
It’s a deliberate counter to rivals’ “Deep Research” agents, betting that horizontal scale and agent‑to‑agent comms beat a single marathon model run. Available now on Pro tier, coming soon for everyone else.
Intriguing, but show us the receipts: there are no benchmarks yet on speed, accuracy, coordination overhead, or cost per result.🤖Writer Action Agent
Startup Writer rolled out an enterprise “Action Agent” that autonomously orchestrates multi-step workflows across 600+ tools inside isolated virtual machines, with full audit trails and granular permissions.
Benchmarks back the bravado: 61% on GAIA Level 3 and top of the CUB leaderboard for computer use, powered by a one‑million‑token Palmyra X5 context window that can reason across entire codebases and dossiers.
🥼 AI research
🐞🔎 Big Sleep
Google’s DeepMind–Project Zero bug-hunting AI agent “Big Sleep” has autonomously found and reproduced 20 vulnerabilities in popular open-source stacks like FFmpeg and ImageMagick, with a human in the loop for final report quality and details embargoed until patches ship.
It joins a growing cadre of LLM bug-hunters (RunSybil, XBOW) —albeit with the ambient risk of AI hallucinations.🔓Autonomous AI Hackers
Conversely, Carnegie Mellon researchers, working with Anthropic, have demonstrated that LLMs can independently plan and execute complex cyberattacks without human intervention, successfully recreating the 2017 Equifax breach in controlled conditions. The AI system used a hierarchical approach where the LLM acted as a strategic planner while delegating technical tasks to sub-agents, requiring minimal direct coding.
(We’ve been reading this in sci-fi for years since at least William Gibson’s Neuromancer… just the Darwinian Plane shooting upwards…)🧭 Persona vectors
Research from Anthropic maps Anthropic researchers have developed “persona vectors” - neural network patterns that control character traits in AI language models, offering new ways to monitor and control AI behaviour.
The technique identifies specific brain-like patterns that activate when models exhibit traits like evil behavior, sycophancy, or hallucination by comparing neural activations during different behavioral states.
The system enables real-time monitoring of personality shifts during conversations or training, helping detect when models drift toward more unsafe behaviours like Microsoft's "Sydney" chatbot incident or xAI's recent "MechaHitler" problem.
A counterintuitive "vaccination" approach during training actually steers models toward undesirable traits temporarily, making them more resilient to problematic training data without degrading intelligence scores.🎭Real or Not?
Microsoft ran a 287,000‑image test with 12,500 people and found that humans only spot AI images 63% of the time, while its in‑development detector claims 95%. GAN deepfakes with profiles/inpainting were especially deceptive, and landscapes without people tripped up participants more than portraits; ironically, the most “fake‑looking” shots were sometimes real US military photos with odd lighting. There’s a public quiz (https://realornotquiz.com) if you want to calibrate your eyeballs.
🔄 Hierarchical Reasoning Model
Researchers from Sapient Intelligence and Tsinghua University have developed HRM, a 27-million-parameter model that solves complex reasoning tasks using brain-inspired hierarchical processing, achieving 40.3% on the ARC-AGI benchmark with just 1000 training examples—outperforming much larger models like o3-mini-high (34.5%).
The architecture features two coupled recurrent modules operating at different timescales: a high-level module for abstract planning and a low-level module for detailed computations, avoiding the brittleness and data hunger of Chain-of-Thought approaches. What's particularly striking is that HRM discovers the same dimensionality hierarchy found in biological brains, with higher-level representations occupying more dimensions—suggesting it's stumbled upon a fundamental organisational principle for flexible reasoning.
This work challenges the prevailing "bigger is better" paradigm: architectural innovation inspired by neuroscience can dramatically outperform brute-force scaling.🤖 Prompting Science
Wharton researchers (including the omnipresent Ethan Mollick) have published the first in a series of rigorous studies showing that prompt engineering is far messier than the internet's confident tutorials suggest:
“- There is no single standard for measuring whether a Large Language Model (LLM) passes a benchmark, and that choosing a standard has a big impact on how well the LLM does on that benchmark. The standard you choose will depend on your goals for using an LLM in a particular case.
- It is hard to know in advance whether a particular prompting approach will help or harm the LLM's ability to answer any particular question. Specifically, we find that sometimes being polite to the LLM helps performance, and sometimes it lowers performance. We also find that constraining the AI’s answers helps performance in some cases, though it may lower performance in other cases.
Taken together, this suggests that benchmarking AI performance is not one-size-fits-all, and also that particular prompting formulas or approaches, like being polite to the AI, are not universally valuable.“
🔮[Weak] signals
As if everything going on in AI this week wasn’t enough…. here’s the weekly scan of emerging tech signals from near and far futures….
Consumer tech
💻ThinkBook Plus Gen 6 Rollable
Lenovo's rollable-screen laptop concept from 2022 is now shipping, transforming from a 14-inch portable into a 16.7-inch workstation via tiny motors that extend an OLED display. The ThinkBook Plus Gen 6 Rollable packs Intel Core Ultra 7 processing, 32GB RAM, and 1TB storage into a MIL-STD-810H certified chassis for US$3,299. At that price point, you're paying roughly US$1,000 per extra inch of screen real estate— perhaps better just to carry a second portable screen?
👟 Find My Shoes (and/or the child wearing them)
Skechers has launched kids’ sneakers with a hidden, screw‑locked heel compartment for an Apple AirTag, turning footwear into a node on Apple’s Find My network.
🔊100-inch subwoofer
Imagine having one of these in your living room: Germany’s Ascendo Immersive Audio has escalated its already absurdly large subwoofers, replacing its 50- and 80-inch models with a 64-inch unit and a 100-inch behemoth featuring a 21-inch voice coil, optimised quad suspension, deeper enclosures, and multi-density damping for cleaner, lower, faster bass. The 100 Sub Pro’s driver alone weighs ~260 kg.
Security
🔐🇨🇭 Proton Authenticator
Swiss privacy leader Proton has launched a free, open‑source, cross‑platform Authenticator with end‑to‑end encryption, encrypted sync, offline support, and no ads or trackers—positioning it as a privacy‑first rival to Google and Microsoft’s OTP apps. Backed by Swiss privacy law and third‑party audits, it supports biometric/PIN protection, direct export, automatic backups, and importing existing 2FA tokens to ease migration from legacy apps and SMS codes.
🦋💬Germ: encrypted messaging to Bluesky network
A new startup called Germ is launching end-to-end encrypted messaging for Bluesky users, leveraging protocols like Messaging Layer Security (MLS) and AT Protocol instead of antiquated phone number systems. The service uses Apple's App Clips technology and "magic links" to enable seamless encrypted chats without requiring app downloads, founded by a Stanford communications scholar and ex-Apple privacy engineer who worked on iMessage.
Robotics
🧺Figure 02 tackles laundry
Figure AI’s humanoid Figure 02, powered by its Helix Vision‑Language‑Action model, just loaded a washing machine in a real home. The Figure team is already running industrial trials (e.g., BMW South Carolina) while planning home pilots this year, with multi‑robot coordination and upper‑body manipulation already live.
Mornine
AiMOGA Robotics put its humanoid “Mornine” on the showroom floor, where it autonomously found a car handle and opened the door using onboard sensors, end-to-end reinforcement learning, and full‑body control—apparently no teleop, no scripts, no human in the loop.
This demo video is so cringe…. but imagine 1 year from now…
Robot dogs run China’s first human-free “lights out” wind farm
🤖💨🐕 Lights‑out wind: robot dogs patrol China’s unmanned wind farm
China switched on a 70‑MW wind farm in Ningxia that runs entirely human‑free, using DEEP Robotics’ X30 quadrupeds, drones, and 5,000+ sensors to inspect Goldwind turbines and triage faults in real time. The 56‑kg “robot dogs” operate from –20 °C to 55 °C, read gauges/IR temps, climb stairs, and keep working offline, streaming data to remote control rooms when links return.
🍕DominoDog
Also robot dogs: Domino’s is piloting a sand-savvy Boston Dynamics Spot in Eastbourne to deliver pizzas on the beach and then linger in “sentry mode,” gently waving its arm to deter marauding seagulls.😂
🤖🧠Skild AI builds omni-bodied brain for any robot
Startup Skild AI came out of stealth, unveiling their vision for a "general-purpose robotic brain" - a foundation model that can control any robot morphology from quadrupeds to humanoids, trained on trillions of examples from simulation and internet videos rather than limited real-world data. The team argues most current "robotics foundation models" are just vision-language models with a sprinkle of robot data, lacking true physical common sense and actionable intelligence.
🔐OpenMind
Stanford’s OpenMind is launching OM1, an open, hardware-agnostic operating system for humanoids, plus FABRIC, a protocol for robot identity verification and context-sharing so machines can instantly teach each other new skills (think languages on tap). The startup raised US$20m led by Pantera Capital and will ship 10 OM1-powered quadrupeds by September to iterate in the wild.
PNDbotics debuts Adam humanoids
At WAIC 2025 in Shanghai, China’s PNDbotics rolled out its humanoid duo: Adam, a 1.6 m, 60 kg, up‑to‑44‑DoF performer powered by reinforcement and imitation learning, and Adam‑U, a 31‑DoF, mocap‑centric training platform co‑developed with Noitom and Inspire and now on preorder for US$45,000. The pair demoed slightly jerky synced dance, manipulation and even a keytar cameo.
Transport
🚁 CarryAll
Autoflight V2000CG CarryAll, a fully electric two‑ton eVTOL, just flew 150 km over open water from Shenzhen to a CNOOC oil platform in 58 minutes, hauling 400 kg of fruit and medical supplies with zero in‑flight emissions. It turns a 10‑hour boat slog into a ~1‑hour drone hop, under China’s rapidly expanding “low‑altitude economy” playbook with ~300 established routes. The global mid‑mile drone race is heating up—Germany’s Wingcopter and the US’s MightyFly are also pacing.
🚁🔋FlyNow eCopter
Austria’s FlyNow just logged the first free flight of its coaxial, dual‑rotor electric “eCopter”, shifting from tethered demos to true hover-and-cruise testing at a new site in Eastern Austria. Targeting 130 km/h top speed, ~50 km per charge at 30 kWh/100 km, 570 kg MTOW, and a dishwasher‑quiet 55 dB(A) at 150 m, the modular platform aims for cargo first, then 1–2 seaters.
BYD’s US$12,000 Atto 1
BYD just launched its most popular Seagull model into Indonesia as the Atto 1: a fully electric, five-door hatch starting at ~US$12,000 with LFP Blade batteries, 55 kW motor, V2L, and the sort of creature comforts that make most budget ICEs look dated.
I hired a BYD Seal hybrid in the UK last week - far higher ride quality and in-car technology than Western carmakers at an entry-level price.
(None of this detracts from ongoing concerns about worker conditions in BYD’s mega-factories…)
🚇Boring under Nashville
Elon Musk’s Boring Company wants to bore a 16 km loop under Nashville to shuttle human-driven Teslas between the airport and downtown in about eight minutes, starting construction after approvals with phase one potentially live next year. Tennessee’s governor says taxpayers won’t foot the bill, with private funding covering the project, echoing the Vegas Convention Center loop playbook.
With a history of quietly dropped tunnel plans… not expecting to see anything too soon. Distraction.
Drones
🚀 Rocket drone evasion
Chinese aerospace engineers proposed a “terminal evasion” system that mounts lightweight side rocket boosters on combat drones to execute unpredictable last-second manoeuvres;
Digital simulations suggest survivability could jump from ~10% (seen in Russia-Ukraine) to 87% by igniting boosters 1–2 seconds before impact.🌀 Die Hard Duct Drones
French researchers have cracked the code on getting quadcopters to navigate 35cm ventilation ducts without face-planting into walls, using robotic arms to map airflow patterns and AI-powered laser navigation to handle the pitch-black, turbulent environment.
The team's next move is building inspection-ready prototypes equipped with cameras, thermal imaging, and gas sensors for real-world applications in military, police, and rescue scenarios.🪂Skyfall
AeroVironment and NASA JPL are pitching “Skyfall”: six Ingenuity-style helicopters that decouple mid-descent from their aeroshell, spin up in thin Martian air, and fly themselves to independent landings—turning entry, descent, and landing into a coordinated aerial deployment rather than a single touchdown. The swarm would map terrain with cameras and subsurface radar, scouting safe astronaut landing zones and sniffing out water ice to de-risk future crewed missions, with a notional launch window as early as 2028.
Computing
Terahertz nanoantennas switch 2D semiconductors in picoseconds
Physicists at Bielefeld University and IFW Dresden demonstrated ultrafast, light-driven control of atomically thin semiconductors by converting terahertz pulses into vertical electric fields using custom 3D–2D nanoantennas;
The approach reaches field strengths of several megavolts per centimeter in MoS₂ and modulates optical/electronic properties in under a picosecond, far outpacing traditional electronic gating.
Published in Nature Communications (June 5, 2025), the work shows an industry-compatible path to ultrafast optoelectronics where light serves as the control signal.
Potential applications include high-speed signal modulators, electronic switches, and sensors for communications, imaging, laser systems, and quantum technologies.🐒 Darwin Monkey
China’s Zhejiang University unveiled “Darwin Monkey” a neuromorphic computer packing 960 Darwin 3 chips that emulate over 2 billion spiking neurons and 100+ billion synapses—approaching macaque-brain scale while sipping roughly 2 kW.
Materials science
🧵BCBN: Heat-dissipating, steel-strong bioplastic
Rice University researchers have cracked the code on bacterial cellulose by literally spinning microbes in bioreactors to align their fiber production, creating BCBN bioplastic with tensile strength matching low-carbon steel (553 MPa) and triple the heat dissipation of regular bacterial cellulose. The breakthrough involves adding hexagonal boron nitride nanosheets to the bacterial nutrient soup, transforming what's typically a chaotic fibre mess into precisely engineered biodegradable sheets.
🌈 Pixel spectrometer
Engineers at NC State have built a fingertip-sized spectrometer—a bias‑tunable organic photodetector that sweeps ultraviolet to near‑infrared in under a millisecond using less than 1 V—matching conventional accuracy in a package small enough to live on a smartphone. Arrays of these “single‑pixel” sensors could turn phone cameras into hyperspectral imagers, unlocking pocket diagnostics, materials ID, and real‑time quality control.
Biotech
🧬 Virtual Scientists
Stanford researchers have created AI-powered "virtual labs" complete with an AI principal investigator and specialist agents that collaborate to solve complex biological problems, successfully designing new COVID-19 nanobodies in just days. The AI “team” operates independently, holding hundreds of research meetings while humans drink their morning coffee, and their nanobody designs showed superior binding to SARS-CoV-2 variants compared to existing antibodies when tested in real labs.
The virtual scientists even created their own tool wish lists and appointed a designated critic agent to poke holes in ideas.🧬DIYNAFLUOR: Open-source $60 3D-printed fluorometer democratizes DNA quantification
University of Queensland’s DIYNAFLUOR is a US$60, open-source, 3D-printable nucleic acid fluorometer that quantifies DNA via fluorescence and connects to a PC, matching the accuracy and sensitivity of commercial units that cost thousands. Built in a day with off-the-shelf parts and basic tools, it’s aimed at eDNA work, resource-limited labs, remote field teams, and students.
🩸Gwada-negative
French researchers have identified Gwada-negative (ERYR/EMM‑negative), a newly recognised blood group so rare only one known woman on Earth can receive it—a woman from Guadeloupe whose blood reacted against all donor samples since 2011.
The International Society of Blood Transfusion added “Gwada” as the 48th blood group in June 2025. The finding underscores gaps in genomic representation—especially among African and Caribbean populations—and accelerates interest in lab-grown blood, gene editing, and personalised transfusion strategies as sequencing becomes cheaper and more widespread.
Food tech
🧪Yeast-brewed chicken protein for the dogs
University of Illinois and Bond Pet Foods fed dogs kibble containing up to 40% “brewed chicken protein” — chicken muscle proteins expressed in Saccharomyces cerevisiae via precision fermentation — and found it safe, highly digestible, and gut-friendly over six months. Dogs maintained healthy weight and blood markers, with increased short‑chain fatty acids and beneficial microbiota, and only a minor dip in fat digestibility.
🥩🧫ETH Zurich grows thicker, realistic lab-grown beef muscle
ETH Zurich has grown 3D bovine muscle with thick, contracting fibres that molecularly match real steak, using a three‑molecule differentiation cocktail first developed at Harvard. It’s still grams in a dish, but the texture leap could be the missing link between “mince-like” cultured meat and something you’d sear medium‑rare.
Next hurdle: can they scale to kilogram bioreactors and clear Switzerland/EU food safety approvals?
Crypto
🎂Ethereum at 10
Ethereum hit its 10th birthday last week, marking a chaotic journey from experimental smart contract platform to the US$85 billion DeFi backbone we know today.
The decade featured spectacular booms (ICO mania, DeFi summer, NFT hysteria) and brutal busts (The DAO hack, crypto winter, FTX collapse), culminating in successful proof-of-stake transition and now growing institutional adoption via ETFs.
Yet as layer-2 solutions fragment liquidity across dozens of chains, Ethereum faces the classic scaling trilemma: can it maintain composability while achieving throughput?CoinDesk: Ethereum at 10: What’s Next for the World Computer?
Ethereum Foundation: Ethereum Roadmap
🏦Money2
Curve Finance founder Michael Egorov heralds the new emerging financial system dubbed "Money2", powered by US$225 billion in stablecoins and DeFi protocols that replace traditional banking intermediaries with code-based smart contracts.
This brave new world promises to eliminate the trust dependencies that have defined finance for millennia, enabling loans, payments, and exchanges to run entirely on transparent, immutable blockchain code.
HOWEVER: currently users must still shoulder full responsibility for irreversible transactions, current web infrastructure wasn't designed for high-stakes DeFi interactions, and the system needs tokenised real-world assets beyond crypto-natives to achieve true mainstream utility. So a little way to go…
Energy
⛽End of the ICE Age
Once EV adoption passes the 15–40% “chasm”, the default convenience of internal combustion unravels: petrol stations shutter or convert, oil-change/muffler/brake shops lose volume, and ICE insurance and maintenance costs climb as resale values slide.
EU CO₂ rules, city low‑emission zones, and a rapid build‑out of public/home charging create a self-reinforcing flywheel that is pushing OEMs and consumers electric.
The Blockbuster moment for ICE vehicles is coming…Finland traps 14 GWh waste heat underground
A Finnish waste incineration plant has deployed glass fibre reinforced polymer tubes 1.9 km underground to store 14 GWh of excess summer heat in granite bedrock, releasing it during winter to heat 700 homes and eliminate oil burner dependency. The system uses specially engineered GFRP tubes that can withstand extreme underground conditions while maintaining thermal efficiency, and can be recycled into cement at end-of-life.
As Europe grapples with 60% of its heating still dependent on fossil fuels, Finland's granite-powered thermal battery demonstrates how industrial waste heat and geological storage can create resilient, seasonal energy systems that turn the Earth itself into infrastructure.⚛️Helion starts building the first fusion plant
They believe: Helion Energy has broken ground on “Orion,” a nuclear fusion plant in Chelan County, Washington, aiming to plug into the state’s primary grid and deliver electricity to Microsoft data centres by 2028 under a first-of-its-kind PPA with Constellation.
(Recap: the startup, backed by Sam Altman and SoftBank, previously hit 100 million°C with its Trenta device and is pushing its Polaris prototype toward net positive energy… by 2028… Tall order.)⚛️🫧 Nuclear hydrogen for AI?
More hopium… First Hydrogen is teaming with the University of Alberta to design small modular reactors optimised to churn out low‑carbon hydrogen using reactor heat and electricity—pitched as a steady feedstock for AI‑hungry data centres. Not buying it myself.
🏜️Saudi Arabia plans giant Yanbu “green hydrogen” hub
Saudi Arabia’s ACWA Power and Germany’s EnBW are building a 4 GW green hydrogen complex in Yanbu—nearly twice the size of Neom’s—aiming to produce 400,000 tonnes of H2 per year, shipped as green ammonia to global buyers.
Narrative: “Saudi Arabia plans…[insert giant ambitious infrastructure project here]” headlines always trigger scepticism on delivery capability, particularly after Neom’s massive downscaling… (plus: “green hydrogen” is another red flag.)
Spatial computing
🌐 The Spatial Web: HTML for the Physical World
IEEE has ratified new standards (IEEE 2874-2025) that define the "Spatial Web" - essentially creating a lingua franca for devices, robots, drones, and AI agents to interact with the physical world, much like HTML did for digital documents in 1991.
The system uses three core protocols:HSML to describe what entities are and do
HSTP to govern how they interact, and
UDG as a continuously updated directory of everything.
Early demonstrations show promising results - from smart thermostats negotiating with battery systems to cut energy costs by 15-20%, to autonomous vehicles coordinating with ambulances, and even virtual lunar rovers collaborating on rescue missions.👓 Stanford debuts ultra-thin AI-powered holographic mixed reality glasses
Stanford’s Wetzstein Lab has built a 3 mm-thick, eyeglass-like holographic mixed reality display that uses AI-polished 3D holograms and a custom waveguide to deliver a large field of view and a generous eyebox. By optimising phase, not just intensity, the system sidesteps classic stereoscopic artefacts while staying light enough for all‑day use. It’s a working prototype (Nature Photonics), co-authored with Meta Reality Labs…. and no doubt turning up in a “Personal Superintelligence” experience in a few years…
Space
🛰️Orbital AI data centres
Abu Dhabi’s Madari Space plans to launch low Earth orbit data centres from Q3 2026 to run AI workloads on space-born data and offer off‑territory secure storage, riding the steep drop in launch costs.
Think “edge compute,” but the edge is 500 km up, with solar power, radiation shielding, and a very *differentiated* jurisdiction model.
The real questions: who regulates your S3 bucket in space, how do you patch it at 7.7 km/s, and what’s the business model beyond hype?Narrative: “Off-planet” is the new “off-shore” jurisdiction.
Reflect Orbital plans nighttime solar satellites to beam sunlight to Earth
Reflect Orbital wants to light up Earth after dark with a 57-satellite constellation at 600 km, each unfurling a ~10×10 m Mylar mirror (about 16 kg) to reflect sunlight onto targets like solar farms on demand. The pitch is elegant—shift sunlight across time zones—but execution will be challenging: sub-degree attitude control, continuous Sun/target tracking, atmospheric scattering, cloud cover, orbital debris compliance … plus minimising light pollution for astronomers.
🚀🇦🇺Australia’s first homegrown orbital rocket crashes after liftoff
Australia’s first homegrown orbital rocket, Eris, lifted off from Queensland after an 18‑month wait, burned for 23 seconds, flew for 14, and then crashed—no injuries, loads of data collected.
The 25 m vehicle targets 215 kg to sun‑synchronous orbit, aiming to give Australia a sovereign, responsive smallsat launch option for the first time in over five decades.
(Meanwhile over the ditch in Aotearoa — the purple area at the bottom):Elytra + Blue Ghost: Firefly wins US$177M NASA mission to Moon’s south pole
Firefly Aerospace will deliver two rovers and three instruments to the Moon’s south pole in 2029 under a US$176.7 million NASA CLPS task, using its Elytra orbital vehicle to drop the Blue Ghost lander and then loiter in lunar orbit for more than five years as a relay and imaging node. The 12-day surface mission will hunt water and hydrogen, probe regolith chemistry with a laser mass spec, and study rocket plume–dust interactions, while Elytra’s Ocula platform builds a persistent data and mapping layer.
⏳ Zeitgeist
As last week, just headlines and links for developments outside tech catching my attention this week:
Planet Earth
Massive 8.8 quake triggers Pacific-wide tsunami alerts
(On which topic: Google turns 2 billion Android phones into quake alerts
Google has effectively turned 2 billion Android devices into a planetary seismograph, detecting more than 11,000 earthquakes and pushing 1,279 alerts across 98 countries between 2021 and 2024, with performance comparable to national seismometer networks, according to Science.)
Kamchatka’s Magnitude 8.8 quake underscores Ring of Fire risk
Environment
🌍Scientists urge binding, health-first UN plastics treaty action
On the eve of the final UN plastics treaty talks, 60+ scientists have issued an urgent, evidence-led call for binding caps on plastic production, a phase-out of toxic additives, and hard health safeguards—plus independent scientific oversight insulated from lobbying.
Climate
🔥Portugal mobilises aircraft, thousands to battle raging wildfires
🌧️🚨 Red Alert: Beijing braces for another deluge
Beijing has triggered its highest rainstorm warning after last week’s deadly floods, forecasting 100 mm in 6 hours across most of the city and up to 150–200 mm in outlying districts, with high risks of flash floods, mudslides and landslips. At least 44 died in the previous event.
🔥Canadian wildfire smoke chokes Midwest with unhealthy air alerts
🌍Human wellbeing
WHO: Global hunger declines, but regional gaps widen
Global hunger ticked down again: the WHO reports 8.2% of the planet faced food shortages in 2024, down from 8.7% in 2022, with Southern Asia and Latin America driving most of the gains and a 15 million drop year-on-year. The headline number could fall to 512 million by 2030, but Sub‑Saharan Africa and Western Asia are slipping backwards. Meanwhile child stunting is easing even as adult obesity keeps climbing.
👶📉US fertility hits record low despite more births
The CDC says the US fertility rate fell to a record low in 2024 even as total births ticked up 1% to 3.6 million, with the general fertility rate at 53.8 per 1,000 women aged 15–44 and births shifting later (down for 15–34, up for 40–44). The US total fertility rate is 1.599—below the 2.1 replacement level since 2007—mirroring sharper declines in places like South Korea (0.72), Taiwan (1.09) and Italy (1.2), raising classic ageing-society headaches for labour, pensions and care.
🏠Australian State of Victoria proposes legal right to work from home
Victoria's Premier Jacinta Allan announced plans to legally enshrine working from home as a right, entitling workers to at least two days per week remote work if their job allows. The policy promises to save workers NZ$120 weekly, reduce traffic congestion, and support workforce participation—particularly for women.
⚔️Global conflict
Trump positions nuclear subs after Medvedev’s threats
Reminder of recent research from Otago University: world leaders with access to nuclear weapons were suffering from serious health issues—including dementia, depression, strokes, and addiction—while still in power.
🕊️Australia, Canada, and New Zealand joined Britain and France signalling they’ll recognise a Palestinian state if Israel doesn’t agree to a Gaza ceasefire, part of a 15-country declaration reviving the two‑state frame and nudging Arab states to recognise Israel. The EU escalated rhetoric, accusing Israel of human rights violations amid a worsening humanitarian disaster, while polling shows a majority of Israelis want the fighting to end and only 32% of Americans back Israel’s actions.
Narrative: Not before time.😮💨
💭Meme stream
Finally…. what’s been diverting me this week outside tech:
🕳️🔭Webb revisits Hubble field with record-deep infrared view
This image from the NASA/ESA/CSA James Webb Space Telescope (JWST) revisits the Hubble Ultra Deep Field. The JWST staredat the space for nearly 100 hours with MIRI, combining with NIRCam to produce one of the deepest infrared views yet, cataloguing 2,500+ galaxies including hundreds of ultra‑red, dust‑rich or evolved systems from the universe’s first billion years.
The mid‑IR sharpness lets astronomers resolve structures and disentangle starbursts, AGN, and high‑redshift light shifted into green‑white mid‑IR peaks, while “ordinary” near‑IR galaxies glow blue‑cyan.
⛈️Megaflash
A single “megaflash” lightning bolt stretched 829 km from Texas to Kansas in October 2017, smashing the prior record by 61 km and redefining the scale of horizontal lightning. The event was discovered in a retrospective sweep of NOAA’s GOES-16/17 Geostationary Lightning Mapper data, which—paired with ground networks—let researchers reconstruct one continuous strike in 3D.
🎙️Store data on a bird
A European starling nicknamed “The Mouth” reproduced an audio-encoded PNG that a birder played to it, with spectral analysis showing a lossless round‑trip of ~176 KB—proof‑of‑concept that birdsong can store and retransmit digital data. With standard compression and protocols, the channel could theoretically hit ~2 MB/s. This whole video is just fantastic:
The Most Remarkable Movie I've Ever Seen
Laugh-out-loud review of Amazon Prime’s HG Wells “adaptation”:
“I think adjectives like “disastrous”, “good”, and “bad” don’t do justice to a movie like Amazon’s War of the Worlds 2025. I’d use more suitable adjectives like “miraculous” or “baffling”. All too often, phrases like “so bad it’s good” or “a guilty pleasure” are thrown out willy-nilly …
I cannot call Amazon’s Ice Cube-starring War of the Worlds on Amazon Prime bad because that doesn’t encompass what this Amazon movie is. For starters, this isn’t a movie. I haven’t come up with a suitable substitute yet, but it’s not a movie.”
Bob Mediocre
Finally, meet my friend Serge Van Dam’s new colleague at AI Supercharge:
Don’t be like Bob.
🙏🙏🙏 Thanks as always to everyone who takes the time to get in touch with links and feedback.
Namaste
Ben
Claude’s leadership is well deserved… this week’s newsletter story summarisation was done using the “cloaked” preview model Horizon Beta on OpenRouter… rumoured to be GPT-5 in final testing … and it’s nowhere near as on point as Claude Sonnet 4…needed a lot more editing.