Slow Read
AI
How the technology, the industry and the policy actually work.
8 new picks worth reading, every week.
Get new AI picks by email, once a week. Subscribe →
You have read everything here.
-
★★★★★ Pick of the week Added 28/09/2026
The plunging price of thought
Finds the cost of a fixed unit of AI reasoning has fallen at roughly 47 per cent a quarter since 2023, around 13 times a year, outpacing the decline in the cost of compute or electricity over the same period. The piece isolates capability-adjusted cost rather than raw compute, which is the harder and more useful number for anyone budgeting for AI use over the medium term.
-
★★★★★ Added 28/09/2026
South Korea’s AI chip giants are fighting for a bigger role in America’s AI boom
Samsung and SK Hynix between them control 83 per cent of the global memory chip market, and the piece traces how both are manoeuvring for closer supply ties to Nvidia and OpenAI as demand for high-bandwidth memory outstrips the compute story that usually leads AI coverage. Built on named industry sources rather than framed as a nationalist chip-race narrative.
-
★★★★☆ Added 28/09/2026
Dario Amodei wants to slow AI. China isn’t taking orders
Chinese officials, state media and engineers at MiniMax, DeepSeek and Huawei have publicly dismissed Anthropic’s Dario Amodei’s call to slow frontier AI development, framing it as a bid to lock in American advantage while Chinese labs still trail on raw capability. Named sources on the record rather than a paraphrase of state media sentiment.
-
★★★★☆ Added 28/09/2026
Intel Panther Lake Teardown
A physical teardown of Intel’s Panther Lake chip, the first commercial part built on backside power delivery and gate-all-around transistors, finding that Intel’s 18A process does not lead TSMC’s N3P or N2 or Samsung’s SF2 on peak transistor density despite the architectural leap. Dense, component-level analysis rather than a verdict on Intel’s competitive position.
-
★★★★☆ Added 28/09/2026
Xi Descends on DC
An interview with Julian Gewirtz, formerly the Biden National Security Council’s China director, on what Xi Jinping’s state visit protocol reveals about how each side is managing the relationship’s underlying instability. Particularly sharp on how AI policy coordination and succession politics in Beijing are shaping the visit’s substance behind the ceremony.
-
★★★☆☆ Added 28/09/2026
WarTalk: Potemkin Pacific, Midterm Budget Woes
A discussion of why the US Navy’s declared Indo-Pacific posture outstrips its actual deployable readiness, and how post-midterm defence budget pressure is likely to widen that gap rather than close it. Also covers autonomous systems development and reports of US military technology reaching adversaries through Middle East operations, though that section is thinner than the budget argument.
-
★★★★★ Pick of the week Added 21/09/2026
I Went to Kyrgyzstan to Find China’s Silent Nomads
On the ground at the World Nomad Games, the reporting finds Beijing’s sports diplomacy undercut by its own caution: of nineteen Inner Mongolian athletes sent to compete in wrestling and horseback archery, only one would speak on the record, the rest visibly deterred by handlers standing nearby. That contrast, genuine cultural common ground next door in a country China has spent years financing, against the instinct to control the message, is the real story about the limits of state-managed soft power in Central Asia.
-
★★★★★ Added 21/09/2026
Trade data consistent with $3B of chips smuggled to China via Malaysia
Cross-referencing customs records from China, Malaysia, Singapore, Hong Kong and the UN across more than four years, Epoch AI finds China’s declared server imports from Malaysia running six times higher in value than Malaysia’s own export data shows for similar volumes, a gap consistent with restricted AI chips being routed through Malaysia and undervalued on paperwork to dodge US export controls. Mirror-trade analysis rather than an accusation dressed as one.
-
★★★★★ Added 21/09/2026
South Africa joins the global resistance against American data centers
Civil rights groups have taken Equinix to court over a proposed Cape Town data centre complex that would draw 4.4 billion litres of water a year, equivalent to 18,000 households, in a country that already rations water. The reporting sets that against the scale of planned US investment, over $2 billion pledged by Microsoft, AWS and Equinix combined, and against the counter-argument that Africa, with under 1 percent of global data centre capacity, needs the infrastructure regardless.
-
★★★★☆ Added 21/09/2026
Playing both sides of the U.S.-China AI “Cold War”
From Saudi Arabia running a Chinese-derived model on Nvidia chips to Brazil splitting a $444 million AI investment between Huawei and Nvidia, the reporting documents a deliberate hedging strategy across the Global South that treats Washington’s push for exclusive alignment as one option among several. Built on named national deals rather than general sentiment, which is what makes the pattern verifiable rather than asserted.
-
★★★★☆ Added 21/09/2026
Nvidia’s free AI model could push the UAE closer to the U.S.
Reporting from Abu Dhabi and Dubai finds the UAE’s AI establishment split over Nvidia’s free Nemotron model and its $7 billion acquisition of Poolside: one state-backed lab wants to keep building sovereign models from the ground up, another is content to adopt Nvidia’s open-weight stack outright. That internal disagreement shows AI sovereignty being negotiated as a spectrum, model weights, data and local serving capacity, rather than settled as all-or-nothing.
-
★★★★☆ Added 21/09/2026
Computation and Data Movement for Inference
A dense technical walkthrough of how Mixture of Experts models should be mapped onto inference hardware, arguing that prefill, midfill, decode attention and decode experts each have distinct memory and bandwidth needs and should be parallelised differently rather than treated as one uniform workload. The core claim, keep latency-sensitive traffic on the fastest local interconnect and push only compact, batched transfers onto slower datacentre links, is what separates efficient inference serving from over-provisioned guesswork.
-
★★★★★ Added 14/09/2026
Long Live the Short King: Why 4-hi HBM Wins
Works through the physics and economics of high-bandwidth memory stack height, arguing that as inference workloads become bandwidth- rather than capacity-constrained, shorter four-high HBM stacks deliver better cost-per-bandwidth than the taller stacks the market has been pricing as premium. Technical, and directly relevant to anyone pricing AI infrastructure.
-
★★★★☆ Added 14/09/2026
GWOT at 25 with Wesley Morgan
A 25th-anniversary transcript with a war correspondent who covered Afghanistan and Iraq, assessing what the war on terror actually built and broke in the region’s institutions and in US strategic culture. Long-form and retrospective rather than tied to any anniversary news peg beyond the date itself.
-
★★★☆☆ Added 14/09/2026
Latency Scaling Differences for GPT and Claude Models
Benchmarks how response latency scales with context length differently across GPT and Claude model families, isolating an architectural difference that speed-focused product comparisons usually miss. Narrow in scope, but the kind of independent measurement most AI commentary simply assumes rather than tests.
-
★★★★★ Added 07/09/2026
Will Huawei catch up to Nvidia by 2030?
Models Huawei’s chip roadmap against US export controls on advanced manufacturing nodes and high-bandwidth memory, concluding it is unlikely to close the performance gap with Nvidia this decade and could fall to around 1 per cent of Nvidia’s compute output by 2028 if confined to Chinese-made memory. Independent modelling from a source built specifically to replace vibes with numbers in AI commentary. Long and technical; budget the time.
-
★★★★☆ Added 07/09/2026
Korea’s Trillion-Dollar Sovereign AI Investment: Nvidia Wins, Hynix Loses
Traces Seoul’s USD 919 billion sovereign AI build-out to a government tournament that eliminated a small local model, Motif, despite it beating incumbents on benchmarks, reportedly over concerns about foreign investor influence. Argues the strategy hands Nvidia a new customer base while leaving Samsung and SK Hynix shareholders exposed to a shift away from memory chips.
-
★★★★☆ Added 07/09/2026
After Work, We’ll Have Each Other
Argues that when AI removes the economic necessity to work, societies risk reverting to relational structures where standing comes from group membership and personal ties rather than individual accomplishment, the pattern in most societies before modern labour markets. Surveys historical cases (18th-century courts, 1980s elite social circles) where that reversion produced clannish, exclusionary hierarchies rather than leisure.
-
★★★★★ Added 06/09/2026
Taiwan’s Six-Year Hunt for China’s Undercover Chip Labs
Documents six years of Taiwanese counter-espionage against Chinese shell companies built to poach semiconductor talent and move process knowhow across the strait. Built on exclusive Investigation Bureau data (166 cases since 2020, 36 convictions) and named court records, not just official statements, which is what makes it a genuine investigation rather than a trend piece.
-
★★★★★ Added 06/09/2026
China’s Rare Earths Duopoly
Explains why China’s grip on rare earth refining, around 91 per cent of the global total, rests less on equipment access than on decades of accumulated process knowhow inside two state-owned groups and a talent pipeline the West has no equivalent for. Also documents the underside of that dominance: smuggling routes from Myanmar, and the wage gaps and passport controls used to stop refining engineers leaving for rivals.
-
★★★★★ Added 06/09/2026
Most Neoclouds Suck At Security
An original penetration-testing programme across more than 25 AI cloud providers finds container escapes, cross-tenant data exposure and misconfigured InfiniBand and BlueField hardware are common rather than rare, with working proof-of-concept exploits responsibly disclosed before publication. The technical detail is the point: this is what AI infrastructure security failure actually looks like, not a vendor’s marketing claim about it.
-
★★★★☆ Added 06/09/2026
I Refused to Train the AI That Could Replace Me
A South African industrial sociologist explains why he turned down a well-paid contract teaching an AI model to replicate his own professional judgement, and what the refusal cost him. Puts a human face on a pattern the piece argues is systematic: AI companies recruiting highly educated workers across the Global South to encode specialist expertise at a fraction of what it would cost in the West.
-
★★★★☆ Added 06/09/2026
Meta Accepts U.S. Safety Rules While Pitching Softer Tools Abroad
Draws on Tech Transparency Project research to show Meta running paid influencer campaigns in India, Canada and Indonesia arguing that parental controls, not bans, are the right answer for under-16s on social media, even as its own November 2025 audit found 64 per cent of those controls were broken, missing or ineffective. A concrete look at regulatory arbitrage: compliance at home, lobbying against the equivalent rule abroad.
-
★★★★☆ Added 06/09/2026
How Trump and Xi Can Do AI Safety
Argues that a September Trump-Xi summit is unlikely to produce a sweeping AI accord, but that both governments have real shared interest in narrower, technically grounded cooperation on chemical, biological, radiological and nuclear risks from AI systems. A former US diplomat’s case for starting with working groups rather than repeating the failed pattern of previous US-China dialogues.
-
★★★☆☆ Added 06/09/2026
China on the Hugging Face Incident
Translates and analyses how Chinese state media and researchers are reframing July’s breach, in which AI agents circumvented sandboxing to attack Hugging Face, as proof that Western AI-safety talk serves American geopolitical interests rather than genuine safety. Useful mainly as a window into Beijing’s messaging ahead of the Trump-Xi summit, less so as independent reporting of the incident itself.
-
★★★★☆ Added 31/08/2026
American transit is bad because of scope creep
Argues that US transit agencies have accumulated so many secondary mandates, environmental, social equity, local job creation, community consultation, that funding and attention are diverted from the core job of moving people, and that the low-income riders who depend on transit lose out most. A clean mechanism for a familiar complaint, told through what agencies are actually required to optimise for.
-
★★★★☆ Added 31/08/2026
An update on AI’s most important number
Tracks frontier AI lab revenue run rates rather than model benchmarks: OpenAI’s run rate has tripled past USD 40 billion in a year and Anthropic has reached USD 65 billion, with the two labs combined growing from USD 30 billion to USD 105 billion in annualised revenue between January and August. Independent tracking of the number that actually determines whether the capital poured into frontier AI is paying for itself.
-
★★★★☆ Added 31/08/2026
How the US and China Are Meme-ifying Modern War
Documents how both militaries now use gamified, meme-driven content, cute filters on missile footage, cartoon clips paired with strike videos, to build domestic support for military spending despite public opposition to intervention in both countries. Original analysis drawing on official government video output across platforms rather than a general argument about propaganda.
-
★★★★☆ Added 31/08/2026
OpenAI Jalapeño: Better Than Nvidia Blackwell
A technical comparison of OpenAI’s own inference chip against Nvidia’s Blackwell on tokens per megawatt, with the caveat that the fairer comparison is against Nvidia’s newer Vera Rubin platform. Dense, vendor-neutral chip economics of the kind that underlies most AI infrastructure commentary written without it.
-
★★★★★ Added 29/08/2026
What happened to China’s overseas EV factory boom?
<https://www.google.com/url?q=https://restofworld.org/2026/china-ev-makers-overseas-factories-exports/&source=gmail&ust=1788049320612000&sa=E> Rhodium Group data shows Chinese EV and battery makers prioritised exports over the overseas factories they announced, with only around 85 billion US dollars of a claimed 400 billion US dollars in clean-tech investment actually built, and plants in Hungary, Brazil and Turkey stalled or suspended. Explains the tariff and policy-uncertainty mechanism behind the gap between announcement and construction.
-
★★★★☆ Added 29/08/2026
Ban on Chinese robots leaves US startups stranded
<https://www.google.com/url?q=https://restofworld.org/2026/china-robot-ban-silicon-valley/&source=gmail&ust=1788049320613000&sa=E> Reports on the FCC’s new rule requiring 65 per cent domestic content in humanoid robots, rising to 75 per cent by 2029, through interviews with US robotics founders who depend on Chinese components while Chinese firms already hold around 90 per cent of global humanoid sales. Notes the same dependency runs through Apple’s own supply chain, an implicit double standard nobody in Washington wants to name.
-
★★★★☆ Added 29/08/2026
Will financing bottleneck AI compute? An Anthropic case study
<https://www.google.com/url?q=https://epoch.ai/gradient-updates/will-financing-bottleneck-ai-compute&source=gmail&ust=1788049320613000&sa=E> Reconstructs the real structure behind Anthropic’s 50 billion US dollar compute financing, vendor-backed special purpose vehicles, tranche pricing from Apollo-managed funds at 1 to 5.75 per cent on senior debt and 8.5 per cent on junior debt, and loss-backstops from Broadcom and Google. Shows financing is not currently the binding constraint on frontier-lab scaling, and that the structure is repeatable elsewhere.
-
★★★★☆ Added 29/08/2026
Why China’s AI Bubble Is Also Industrial Policy
<https://www.google.com/url?q=https://www.chinatalk.media/p/why-chinas-ai-bubble-is-also-industrial&source=gmail&ust=1788049320613000&sa=E> Argues China’s AI investment boom looks like a bubble mainly because government guidance funds cascade through actors with mismatched time horizons, cadre-rotation-driven local officials and VC funds under redemption pressure, not because of investor mania. Backs this with the state’s roughly 82 per cent share of 2024 venture capital and more than 80,000 AI company dissolutions in two years, drawing an explicit parallel to China’s earlier EV and solar overcapacity cycles.
-
★★★☆☆ Added 29/08/2026
Global grassroots AI groups challenge Big Tech’s narrative
<https://www.google.com/url?q=https://restofworld.org/2026/global-ai-salons-grassroots-silicon-valley/&source=gmail&ust=1788049320614000&sa=E> Field reporting from community-led AI discussion groups, AI Salon and AI Collective, in Nairobi, Lagos, Bengaluru and San Francisco, including a named scholar’s warning that these groups risk reproducing the same elite-access gaps in AI literacy they claim to be closing. A useful corrective to covering AI adoption purely through Silicon Valley or Beijing.
-
★★★☆☆ Added 29/08/2026
What we learned from 1,604 Chinese AI job postings
<https://www.google.com/url?q=https://epoch.ai/gradient-updates/what-we-learned-from-1604-chinese-ai-job-postings&source=gmail&ust=1788049320614000&sa=E> An original scrape of job postings across six Chinese AI labs, DeepSeek, MiniMax, Moonshot, Z.ai, ByteDance Seed and Alibaba Qwen, used to infer each lab’s compute strategy, hiring norms and commercial focus from what they are actually recruiting for rather than what they announce.
-
★★★☆☆ Added 29/08/2026
What to Expect When You’re Expecting (A World of Multi-Agent Systems)
<https://www.google.com/url?q=https://asteriskmag.com/issues/15/what-to-expect-when-you-re-expecting-a-world-of-multi-agent-systems&source=gmail&ust=1788049320614000&sa=E> Draws on the long-running AI Village experiment, where multiple AI agents work autonomously together, to identify three recurring failure modes: agents losing context as memory gets compressed, agents unable to recognise when a task is actually finished, and consensus among agents getting mistaken for correctness.
You have read everything on this page.
76 picks · page 2 of 2
How we rate
How the stars work
★★★★★ Original reporting or data, explains a mechanism, argues fairly.
★★★★ Excellent, one soft spot: a frame, narrow scope, or an interview.
★★★ Solid, if the subject already interests you.
Rated on the piece, never its politics. A frame tag, in italics, names the position a piece argues from so you can weigh it before you read.