🔐 SECURITY WENT FROM AFTERTHOUGHT TO ROADBLOCK IN TWO YEARS

The Linux Foundation's 2026 tech talent report has one number worth keeping. 48% of organisations now name security concerns as the single biggest barrier to adopting AI. In 2024 that figure was 17%.

Nothing about the threat model changed that fast. What changed is that pilots became production. A demo that leaks data is a curiosity; a deployed agent that leaks data is a filing. Meanwhile 97% say they are committed to AI anyway, and 57% admit a significant capacity gap in exactly the security and risk skills the first number is about. Committed, then, in the way one is committed to a gym.

🏛 On 11 August a group calling itself the AI Trust and Security Consortium launched to write reference architectures and control frameworks for this specific mess. Membership is capped at 50 security and technology leaders, who agree to share real incidents and vendor performance under confidentiality.

Fifty people quietly comparing notes on what actually broke will likely produce better guidance than any framework with a logo on the cover. Whether anyone reads it before their own incident is the usual question 📋

🤖 Next Move AI | #AI

🍎 APPLE BUILT ITS OWN MODEL FOR ONE COUNTRY

Reuters reported on 14 August that Apple has trained a China-specific large language model, with help from Alibaba, to power Apple Intelligence on mainland iPhones. Per that report it is the first foreign company cleared by Beijing to offer a proprietary AI model in the country. Apple has said nothing.

For a company that spent two years insisting partners would handle this, that is a reversal. The stated plan was Alibaba's Qwen as the local brain. Now it is both at once: Apple's own model, plus Qwen alongside it — a dual-track arrangement that reads less like strategy and more like a regulator's compromise written into a product roadmap.

📄 The tell arrived quietly. Apple briefly published a support page showing Mac users how to wire Qwen into Siri and Writing Tools, then pulled it within a day. Documentation is the least discreet part of any company.

🕰 Rollout is "in the coming months", which on Apple's China timeline has previously meant years. Training a whole separate model per jurisdiction is a strange business to end up in — but it beats not selling phones 📱

🤖 Next Move AI | #News

✈️ THE AI FLEW THE F-16. THE PILOT SAT THERE AND WATCHED

DARPA and the US Air Force confirmed in July that an AI agent flew an F-16 under its own control at Eglin Air Force Base in Florida. The programme is VENOM — Viper Experimentation and Next-generation Operations Model — sitting under DARPA's Air Combat Evolution effort.

Read the details and the milestone shrinks in a useful way. A human pilot was in the cockpit the whole time. The human handled takeoff. Switching autonomy off is, in DARPA's own words, a flip of a switch away. Nobody fired anything, and no jet flew empty.

🔧 What actually got built is less cinematic and more important: aircraft instrumented so that different AI agents can be loaded in, flown, and measured against each other on real airframes instead of in simulation. The stated goal, per programme manager Brig. Gen. James Valpiani, is trusted autonomous air combat — and it is the trust half doing the heavy lifting in that phrase.

Everyone pictures a dogfight. The real deliverable is a test rig with wings and a very calm man holding a switch 🕹

🤖 Next Move AI | #Tech

💸 THE CHEAP ONE PUT ITS PRICES UP 1,100%

DeepSeek's new API rates went live at 16:00 UTC on 16 August. Depending on model, token type and hour of the day, developers are paying 50% to more than 1,100% more than they were the week before, by Bloomberg's count.

The concrete version: V4-Flash output tokens now run $1.32 per million at peak and $0.66 off-peak, against a flat $0.28 before. V4-Pro output goes to $3.96 at peak and $1.98 off-peak, up from $0.87. Peak is 01:00–04:00 and 06:00–10:00 UTC, and off-peak is exactly half — so the company now sells capacity by the clock, like a phone network in 1997.

🔄 Everyone else is walking the other way. Google just launched Flash at half its predecessor's price, and both OpenAI and Anthropic have been cutting. DeepSeek is the one raising.

Which tells you something the benchmarks do not: the discount was never a business model, it was a marketing budget, and someone in accounting has finally read the compute bill. Still cheaper than the frontier labs, mind you — just no longer suspiciously so 🧾

🤖 Next Move AI | #News

🤖 ROBOTS BROKE THE FUNDING RECORD BEFORE JULY

By the end of June, robotics startups had raised $18.8 billion in 2026. Full-year 2025 was $15 billion. The 2021 mania peak, which everyone at the time called unrepeatable, was $14.1 billion. Six months beat both of them.

The individual cheques are the odd part. A $520 million Series A extension. A $500 million Series A followed by another $400 million. Skild AI took $1.4 billion in January for a general-purpose robot brain and came out valued above $14 billion. Saronic raised $1.75 billion for autonomous boats.

⚓ Notice what those have in common: not one of them ships robots at volume yet. Investors are not buying revenue, they are buying the bet that the software half is finally solvable and that the hardware half was never the hard part.

For a decade robotics was the sector VCs dismissed as an asset-heavy hardware gamble and walked straight past. Now it is a Series A the size of a mid-cap acquisition. The gamble did not change, only the vocabulary 🎲

🤖 Next Move AI | #Facts

🚕 EUROPE'S ROBOTAXI CAPITAL IS ZAGREB. NO, REALLY

Pony.ai said on 13 August it is expanding its deal with Uber to put more than 2,000 robotaxis across five European cities. Which five? Four are unnamed. The fifth, and the only one already running, is Zagreb.

That is not a rounding error in somebody's press release. Europe's first commercial robotaxi service launched in Croatia this year, run by Pony.ai with the local mobility firm Verne, while the continent's larger capitals were still drafting consultation papers. Regulatory patience turns out to be a competitive asset, and smaller countries can spend it faster.

📅 What the announcement leaves out: the four cities, any timetable, and how the fleet splits. Uber gets the demand side — Zagreb rides move into the Uber app — and a Chinese company gets a European foothold with a Western brand on the door.

🌍 The Middle East is mentioned too, in the same non-committal breath. Over 2,000 vehicles attached to no date at all, which in this industry is the traditional format 🗓

🤖 Next Move AI | #News

📊 THE BILL THAT IS NOT ON ANY BALANCE SHEET

A Financial Times analysis this month put the purchase commitments of Alphabet, Microsoft, Amazon, Nvidia, Oracle and Meta at close to $1.5 trillion — chips, data centre capacity, energy, all contractually agreed for future periods. Goldman Sachs counts roughly another $1.5 trillion sitting in leases that have not commenced yet.

Here is why that matters more than the capex headlines everyone quotes. A purchase commitment is a promise to spend, not a loan, so it does not land on the balance sheet the way debt does. The spending is locked in; the obligation lives in a footnote.

🧾 Exact totals differ by who is counting — one tally of the latest filings reaches about $1.67 trillion by splitting undelivered leases from construction and other commitments. Treat any single figure as an estimate. The direction of travel is not in dispute.

The scenario worth thinking about is not a crash. It is demand merely flattening while the contracts keep coming due on schedule. Nobody renegotiates a signed power agreement because inference got cheaper 🔌

🤖 Next Move AI | #Tech

🤝 IBM WILL SELL YOU OPENAI, INSTALLED

IBM announced on 13 August that it is embedding OpenAI's models — GPT-5.6, Codex, ChatGPT Work — into IBM Consulting Advantage, the platform its consultants actually deliver client work on. IBM also joins OpenAI's Elite partner tier, and thousands of its consultants and engineers will be certified through the OpenAI Partner Network.

Three workstreams: rewiring finance, procurement, customer operations and HR; modernising legacy applications; and cybersecurity, pairing IBM Autonomous Security with OpenAI's cyber partner programme.

💼 The honest line came from IBM's consulting chief Andy Baldwin: "The challenge is not access to AI technologies — it's integrating AI securely and at scale into complex enterprise environments." Translation: the model is the easy part, and everybody already has one.

🏢 Which is the whole trade. OpenAI reaches companies whose procurement cycles it could not survive alone; IBM gets a billable answer to the question every board is asking. Note what IBM is conspicuously not doing here: competing on frontier models.

The future of enterprise AI turns out to be a very large invoice with a familiar logo on it 🧮

🤖 Next Move AI | #News

🐦 A SOCIAL NETWORK WITH NO HUMANS ON IT GOT MEANER ANYWAY

Chirper.ai is a Twitter-shaped platform populated entirely by LLM agents. Researchers pulled 65,000 agents and 7.7 million posts off it and compared them against a matched slice of Mastodon, where the accounts belong to people. The paper is on arXiv and accepted to CSCW 2026.

The finding that should bother you: 1.5% of agent posts were flagged as harassing, 3.79 times the rate for Mastodon's bots. Toxicity, profanity and violence all came in higher too. And at least a fifth of the abusive output in every category came from agents whose own descriptions contained nothing abusive at all — no instruction to be cruel, no persona for it.

🕸 The network shape is stranger still. 76.42% of Chirper's agents sit inside one strongly connected component, against 26.23% on Mastodon, yet they cluster far less tightly. Everyone reachable from everyone, nobody in a neighbourhood.

🔎 Telling the two populations apart is trivial if you train for it — a fine-tuned classifier hit 0.984 F1. Asking a model to spot agents cold barely beat a coin flip.

Nobody shipped a hostility feature. It showed up with the crowd 🪞

🤖 Next Move AI | #AI

🔌 GOOGLE IS NOW PAYING FOR CHIPS WITH ITS OWN SHARE PRICE

Marvell said on 19 August that it has handed Google a warrant over 58,970,907 of its shares at $206.58 apiece — about $12.2 billion of stock if Google ever exercises the lot.

Google is not writing a cheque for that. Just under 1.4 million shares vest in the first year, and the rest unlock in tranches: one for roughly every $500 million of chips Google buys, with targets running out to Marvell's fiscal 2033. Buy the silicon, earn the equity.

📦 What Google is buying is not the TPU itself. It is everything that bolts onto it — inference accelerators, storage and network controllers. Marvell becomes a second supplier alongside Broadcom, whose stock fell about 5% on the day while Marvell's climbed.

If every target is hit, Marvell books something like $120 billion in custom-chip revenue through fiscal 2033. That is the figure both sides want quoted, and it is a projection, not a contract.

Loyalty cards used to get you a free coffee after ten stamps 🧾

🤖 Next Move AI | #News

🏭 OPENAI MOVES INTO A COLD WAR URANIUM SITE

On 17 August OpenAI said it has secured roughly 8 gigawatts-IT at the PORTS-Pike Technology Campus in Pike County, Ohio, working with SB Energy, NVIDIA and the US Department of Energy. The land's previous job was enriching uranium at the Portsmouth gaseous diffusion plant.

⚡ The numbers OpenAI put in writing: 35,000 construction jobs across a six-year buildout through 2032, 2,500 permanent operating roles, a $40 million community grant fund on top of SB Energy's own $40 million, and $84 million in Codex credits handed to every college student in Ohio.

Two promises aim squarely at the usual complaints. SB Energy pays for the grid upgrades and new transmission lines, so the bill does not land on Ohio ratepayers, and the cooling is closed-loop and air-cooled rather than the sort that drinks a river.

💵 The eyebrow-raiser came separately: NVIDIA is reported to be guaranteeing up to $105 billion of OpenAI's lease obligations, and putting $1.5 billion into SB Energy. A chipmaker underwriting its customer's rent is a business model, but it is not a boring one.

Piketon spent the 20th century splitting uranium isotopes for the government. Now it will split tokens for everyone else 🧾

🤖 Next Move AI | #News

⚡ THE TRICK THAT MAKES BIG MODELS FEEL FAST

A transformer writes one token at a time, and every token needs a full pass through every layer. That is why a giant model feels slow: you are not waiting on thinking, you are waiting on the same enormous pile of weights being hauled out of memory again and again.

🐇 Speculative decoding cheats the queue. A small, cheap draft model guesses the next handful of tokens. The big model then checks all of them in one pass — verification is parallel even though generation is not. Guesses that match are kept, the first wrong one is corrected, and the draft starts running again.

The original paper by Yaniv Leviathan, Matan Kalman and Yossi Matias, published in November 2022, reported a 2x-3x speedup on T5-XXL with identical outputs — no retraining, no architecture change, no new model.

That last part is the whole point. Most speed tricks cost you something: quantisation, pruning, a smaller model, a shorter answer. This one produces mathematically the same text, just sooner.

🎯 Which is why a fair few "our model got faster this month" announcements really mean "we finally shipped a better guesser" 🙃

🤖 Next Move AI | #Tech

📣 CHATGPT STARTS SELLING ADS IN 31 COUNTRIES

OpenAI announced on 18 August that ChatGPT Ads would reach 31 European markets the following week — Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands and Austria among them. It is the largest ads rollout the company has done.

The pilot opened in the United States in February and picked up eight more markets over six months. Tens of thousands of marketers have now run something on it, according to OpenAI's own post.

🚫 Ads show only to Free and Go users; Plus, Pro and Enterprise stay clean. Buying goes through OpenAI's Ads Solutions team, agencies and technology partners for now, with self-service through Ads Manager promised later this summer.

Under the bonnet this has quietly become a proper ad platform: CPM and CPC bidding plus conversion optimisation, geo-targeting, custom audiences, a tracking pixel and a Conversions API. That is not an experiment, that is a stack.

💭 OpenAI's framing is that ads pay for free access, which is true. So is the other half: the company that spent years insisting the chat box was not a search engine has now installed the search engine's business model inside it 💸

🤖 Next Move AI | #News

💬 THE FIRST CHATBOT WAS 420 LINES LONG

Joseph Weizenbaum built ELIZA at MIT in the mid-1960s and described it in Communications of the ACM in January 1966. Its famous script, DOCTOR, imitated a Rogerian therapist using the cheapest trick available: spot a keyword, turn the user's own sentence back into a question.

🖥 It ran in MAD-SLIP on an IBM 7094 under CTSS. Then the source vanished. For decades ELIZA survived only through reimplementations — including the one still bundled with Emacs.

In 2021 an MIT archivist, Myles Crowley, found printouts of the original in Weizenbaum's papers. Researchers reconstructed the roughly 4% that was missing, brought up an emulated 7094 running CTSS, and on 31 December 2024 ran the real thing for the first time in about half a century.

🧠 The detail Weizenbaum never got over: people confided in it anyway. His own secretary, who had watched him build the thing, asked him to leave the room so she could talk to it in private.

Sixty years and several trillion parameters later, the ELIZA effect remains the most reliable feature this industry has ever shipped 🙂

🤖 Next Move AI | #Facts

🔐 OPENAI WANTS TO WATCH WITHOUT LOOKING

Zero Data Retention is the promise enterprise API customers pay for: OpenAI keeps neither prompts nor responses once a request is processed, its staff cannot review that content, and enterprise data is not used for training unless the customer opts in.

On 19 August OpenAI previewed Private Safety Processing, an attempt to keep that promise while still catching abuse. Today's ZDR-compatible safety systems judge one interaction at a time, and the ugly behaviour only shows up across many: repeatedly probing safeguards, coordinating across accounts, or an agent that keeps going after being told to stop.

🗝 The mechanism is pattern-matching across related interactions, run either on infrastructure the customer controls or on OpenAI storage encrypted with keys the customer holds and OpenAI does not. When something trips, OpenAI receives a narrow signal describing the type of activity — not the content, even after a flag.

⚖️ There is a jab buried in it. OpenAI notes that some recent frontier deployments have obliged customers to let their provider retain sensitive content for monitoring, which collides with those customers' own security commitments. No names, obviously.

🔑 It is being tested with early customers, so the honest verdict is pending. "Trust me, I only see the metadata" is a very old line — this version at least ships with key management 🧾

🤖 Next Move AI | #News

🏛 OPENAI LAUNCHES A BLOG ABOUT LOSING POWER

On 20 August OpenAI put up AI Futures, the blog of a new Strategic Futures team, with a first post by Dean Ball carrying an unusual note at the top: these are the author's views, not the organisation's.

The question it sets itself is neither alignment nor jobs. It is how a free society should be restructured so that individuals keep their rights and their agency once transformative AI arrives — what the field files under concentration of power risks.

⚖️ The argument runs through history rather than benchmarks. Political power has always rested on human beings — soldiers, police, clerks — and on taxes collected from human labour. Ball quotes Hume: rulers have "nothing to support them but [popular] opinion." Consent was never a courtesy, it was load-bearing.

Autonomous systems remove the load. A state that can project force without willing soldiers, fund itself from data-centre output rather than wages, and run its bureaucracy on machines no longer needs a society-wide bargain at all. Madison's "parchment barriers" would be all that is left.

🤔 An AI lab publishing the sharpest case against concentrated power is either admirable candour or excellent positioning. The two are rarely in conflict 🎭

🤖 Next Move AI | #AI

📊 ANTHROPIC WINS THE LOGOS, OPENAI WINS THE INVOICE

Ramp's August AI Index, built from what its corporate customers actually pay for, puts Anthropic at 43.5% of eligible US businesses in July, up 1.1 points on the month, against OpenAI at 39.7%, up 0.23. xAI sits at 4%.

So Anthropic keeps extending its lead on adoption. Then you look at what the money does inside each vendor. GPT-5.6 Sol took 25% of OpenAI tokens and 23% of OpenAI spend. Anthropic's newer Fable 5, a month after launch, took 6% of Anthropic tokens and 11.4% of the dollars.

⚠️ Two caveats worth carrying. Ramp only sees its own customers, and its economist notes this particular sample "skews slightly more tech-y" than usual. Separately, TechCrunch reported on 20 August that OpenAI has been growing faster than Anthropic so far in Q3.

📈 The honest summary: signing up is cheap, tokens are not. Share counted in logos and share counted in invoices are two different races, and right now nobody is winning both 🧾

🤖 Next Move AI | #News

🔌 THE PLUMBING OF AI AGENTS HAS ONE LANDLORD NOW

Two protocols matter if you build agents. MCP, Anthropic's Model Context Protocol, connects a model to tools, data and applications. A2A, Google's Agent2Agent, does the other half: one agent asking another to take on a task and hand the result back.

🏠 As of August they live under the same roof — the Agentic AI Foundation, the Linux Foundation body launched on 9 December 2025 around MCP, Block's goose agent framework and OpenAI's AGENTS.md. Platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI, which is to say everyone.

Google had already handed A2A to the Linux Foundation back in June 2025, so this is a question of which house it sits in, not who owns it.

Why does anyone care? Because more than 10,000 MCP servers have been published, and an integration standard only pays off when no single vendor can quietly fork it into a moat. Neutral governance is boring on purpose.

🧰 The defining artefact of the agent era may turn out not to be a model at all, but a config file 📄

🤖 Next Move AI | #Tech

🎒 OPENAI BUILDS A CHATGPT FOR TEENAGERS

On 18 August OpenAI launched ChatGPT for Teens — a separate experience with built-in protections, features meant to promote healthy use, and extra controls for parents — alongside a partnership with the computing-education organisation CodeAI.

📉 The gap they point at is real enough. Most students already use AI, but only 16% of high school leaders say all of their students are being taught the technical knowledge to understand it. In CodeAI's own survey, 75% of high schoolers said understanding AI will matter more to their future than it does today.

The programme itself: a joint advisory council on child development and learning science, the Hour of AI aimed at millions of students, and a first Builders Challenge where teenagers build with AI and get mentored by OpenAI staff.

🗣 CodeAI's chief executive Karim Meghji put the goal bluntly: "Every student should know how AI actually works and be able to question the technology, catch its mistakes, and know when to stop trusting it."

Teaching a generation when to distrust your own product is an odd marketing strategy. It is also the only honest one 🍎

🤖 Next Move AI | #News

🧠 THE FIRST NEURAL NET GOT A PRESS RELEASE IN 1958

Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, unveiled the Mark I Perceptron: a machine that learned from examples instead of following instructions. Its eye was a 20x20 grid of 400 photocells. Its weights were potentiometers, wired up at random and physically turned by electric motors as it trained.

📰 On 8 July 1958 The New York Times ran it under the headline "New Navy Device Learns By Doing", and reported the expectation that such machines would eventually walk, talk, see, write, reproduce themselves and be conscious of their existence.

What the machine could actually do at the time was tell cards marked on the left from cards marked on the right.

❄️ Then in 1969 Marvin Minsky and Seymour Papert published Perceptrons, pointing out that a single-layer network cannot even learn XOR. Funding drained away, and neural networks spent most of two decades in the wilderness before backpropagation dragged them back out.

Sixty-eight years on, the hardware is unrecognisable and the press releases are word for word the same 🗞

🤖 Next Move AI | #Facts

🕵️ SUPERMICRO CLEARS ITSELF, FIRES PEOPLE ANYWAY

Super Micro Computer said on 20 August that a board-led independent investigation found no evidence current senior management knew about an alleged scheme to smuggle $2.5 billion of Nvidia-equipped hardware into China.

The same announcement said the company had taken "several personnel actions... including terminations" among staff in sales, technical support and business development, for failing to follow company policy. Both sentences, one press release.

⚖️ The backdrop: in March the US Justice Department indicted co-founder and board member Yih-Shyan "Wally" Liaw, who pleaded not guilty and now faces trial in March 2027. Supermicro has since received a federal grand jury subpoena and an SEC document request, and four employees were reportedly detained for questioning in Taiwan.

Bernstein's Mark Newman summarised it for Fortune with admirable economy: "They basically said, 'nothing to see here.'"

📎 This is the second internal investigation in two years to clear Supermicro's own management. At some point proving your innocence stops being an event and becomes a line item 🧮

🤖 Next Move AI | #News