Tech Digest – March 23, 2026

The Recursive Frontier

China’s MiniMax Confirms Recursive Self-Improvement — Model Handles 30–50% of Its Own Training

MiniMax released M2.7, which the company describes as its “first model deeply participating in its own evolution.” During development, M2.7 updated its own memory, built training infrastructure, and improved its own learning process — handling 30–50% of the reinforcement learning workflow without human intervention, repeated over 100 autonomous cycles. The model scores 56.22% on SWE-Pro, matching frontier competitors, and delivers roughly 30% improvement on internal benchmarks through self-optimisation alone.

Note: Recursive self-improvement was, until recently, a theoretical concern. It is now a product feature shipping from both sides of the Pacific. Institutional timelines that assume stable capability growth rates are built on a fading premise.

Sources: MiniMax, VentureBeat

LLMs Trained to Predict World Events With Superforecaster-Level Accuracy

Mantic and Thinking Machines used reinforcement learning via Tinker to train LLMs on approximately 10,000 geopolitical and current affairs questions — events where resolution was known to researchers but not to the model. The resulting systems are approaching superforecaster-level accuracy on event prediction, with the methodology now scaling to economic indicators and election outcomes.

Note: Strategic foresight in institutions relies on human experts and structured scenario methods. When an LLM matches that accuracy at a fraction of the cost and latency, the question shifts from whether planning departments adopt these tools to how quickly their forecasts become the default input.

Sources: Thinking Machines, VentureBeat

Chip Supply Under Strain

TSMC’s 2nm Booked Through 2028, Nvidia Redesigns Next-Gen Chips, Photonic Alternatives Surge

TSMC’s 2-nanometre process is fully committed through 2028, with annual price hikes expected through 2029. Even Nvidia — TSMC’s largest advanced-node customer at roughly 20% of capacity — cannot secure enough 1.6nm A16 slots for its next-generation Feynman AI platform, forcing a redesign that shifts less critical dies to 3nm. A16 capacity is expected to reach only 20,000 wafers per month by end of 2027. Meanwhile, China’s Huawei-backed Yuanjie Semiconductor, now the world’s second-largest silicon photonics laser chip supplier, has seen its shares surge approximately 780% in a year as AI data centres race to adopt optical interconnects. Founder Zhang Xingang has joined the billionaire ranks on the strength of that demand.

Note: Any digital procurement process that depends on cutting-edge silicon — from cloud contracts to custom hardware — runs through this bottleneck. The photonic alternative is not a curiosity; it is the supply chain diversifying under pressure, with a Chinese company leading the charge.

Sources: TipRanks, Investing.com, TrendForce, Forbes via Orbit RRI

The Office Reorganises

Zuckerberg Is Building an AI Agent to Help Him Be CEO

Mark Zuckerberg is developing a personal AI agent designed to retrieve information, cut through organisational layers, and support executive decision-making at Meta. The project is part of a broader initiative to integrate AI across all employee workflows — on the January earnings call, Zuckerberg declared 2026 the year AI would “dramatically change the way” Meta works. The stated goal: every person inside and outside Meta eventually has their own agent.

Note: The signal is not that a tech CEO uses AI tools. It is that the CEO of a 70,000-person company is publicly modelling a future where organisational layers exist to be compressed by agents. For any institution whose hierarchy is designed to route information upward, this is a preview of the restructuring pressure ahead.

Sources: WSJ, Fortune, Euronews

Snowflake Replaces Entire Writing Team With AI — A Generation Pivots to Trades

Snowflake eliminated its approximately 70-person technical writing team this week, replacing them with an AI system called Project SnowWork that generates API documentation from source code. Labour analysts flagged it as a landmark: one of the first wholesale team replacements by a named AI system at a major public company. Meanwhile, 60% of Americans aged 18–29 now see AI as a threat to their job prospects (Harvard), and Stanford data shows employment in AI-exposed roles for 22–25-year-olds declined 16% between late 2022 and September 2025. The response is measurable: 77% of Gen Z say it matters that their future job is hard to automate, and electrician median weekly earnings — $1,376 — now sit 14% above the national median.

Note: When a $75 billion company names the AI system that replaces an entire department, it stops being anecdotal. The blue-collar pivot is not nostalgia — it is a generation reading the labour market signal before most workforce planning departments have updated their models.

Sources: Business Insider, Benzinga, WSJ, CBS News

Compute Goes Orbital

Blue Origin Files for 51,600 AI Satellites — OpenAI Scales Back Data Centres for IPO

Blue Origin submitted an FCC application on March 19 for “Project Sunrise” — a 51,600-satellite constellation designed for orbital AI computing, distributed across sun-synchronous orbits at 500–1,800 km altitude and powered entirely by solar panels. The company argues that moving AI workloads to space is “the only way to solve the energy crisis” facing the industry. The constellation would complement its 5,408-satellite TeraWave connectivity network. Meanwhile, on Earth, OpenAI has retreated from its most ambitious data centre plans ahead of a potential Q4 2026 IPO, shifting its positioning from builder to purchaser of cloud capacity and targeting roughly $600 billion in total compute spend by 2030 — a figure now explicitly tied to revenue projections. The company recently closed a record $122 billion funding round at an $852 billion valuation.

Note: Two realities in one item. Orbital compute is a serious engineering proposal backed by serious capital — not science fiction. And the largest AI company on Earth just learned that Wall Street disciplines spending in ways that social media does not. Any institution planning digital infrastructure on a 5-year horizon is navigating between both timelines.

Sources: Aviation Week, GeekWire, CNBC

Physical Frontiers

First In Vivo CRISPR CAR T Cell Generation Opens Path to Scalable Cancer Therapy

Researchers have performed the first successful in vivo generation of CAR T cells using CRISPR-Cas9, published in Nature. A dual-vector system targets T cells directly in the body via anti-CD3 antibody-guided delivery, integrating a CAR transgene into the endogenous TCR alpha locus. Tested in humanised mouse models for B cell aplasia and both haematological and solid tumours, the approach could eliminate the costly, time-intensive ex vivo manufacturing process that currently limits CAR T access to a fraction of eligible patients.

Note: Current CAR T therapy costs €300,000–€400,000 per patient and requires weeks of laboratory manufacturing. In vivo generation could eventually turn it into an injection. For public health systems budgeting cancer care, this is the line between experimental and scalable.

Sources: Nature, Innovative Genomics Institute, The Medicine Maker

BYD Brings 5-Minute Flash Charging to Europe on April 8

BYD’s 1,500 kW Flash Charging system charges compatible EVs from 10% to 70% in five minutes, adding roughly 500 km of WLTP range. The Denza Z9 GT — the first vehicle to carry the system — launches in Europe on April 8, 2026. BYD has already deployed thousands of Flash Charging stations across China and is beginning a European rollout, bringing charging speeds that effectively eliminate the refuelling time gap between electric and combustion vehicles.

Note: Any municipality mid-procurement on EV charging infrastructure should note that station throughput assumptions are about to shift. A five-minute charge cycle means a single station can serve dramatically more vehicles per day — and the planning models written around 30-minute sessions need revising.

Sources: Wired, Electrek, InsideEVs

China’s Humanoid Robot Rental Market Explodes — Prices Drop 90% in One Year

Humanoid robots built on Unitree and AgiBot platforms are now commercially rented across China for retail, events, and hospitality. Rental prices have collapsed from 10,000–20,000 yuan per day in spring 2025 to as low as 1,796 yuan on JD.com — a decline exceeding 90%. Over 1,500 robot rental companies registered in China in 2025, a 48% year-on-year increase, with the market projected to reach 10 billion yuan in 2026. Unitree plans to ship 20,000 humanoid robots this year, up from roughly 5,500 in 2025.

Note: The commercialisation curve matters more than the technology curve. When rental prices drop 90% in twelve months, robots cross from demonstration to deployment economics. The EU has no equivalent rental market, no production base at this scale, and no subsidy framework for commercial humanoid deployment.

Sources: People’s Daily, Rest of World, Interesting Engineering, eWeek