majlis@daily — 12 Nov 2026 — Daily AI News — AI Majlis Daily
majlis@daily:~$ cd ~ / articles / 2026-11-12
AI Majlis Daily

Daily AI News — 12 Nov 2026

# Thursday, 12 November 2026 · The AI stories that matter worldwide, ranked
# jump to a story
01OpenAI: frontier models are getting harder to monitor through their reasoning [Research]As models scale, the thoughts they show us explain less of what they actually do.02DeepMind's AlphaGenome 2 reads DNA cell by cell to pinpoint disease variants [Research]A next-generation model for predicting how specific cells interpret the genome.03Google launches the Gemini Decisions API for fast, calibrated AI judgments [Industry]Score up to 1,000 yes/no, choice and rubric questions in a single multimodal pass.04UK AI chip startup Fractile hits $7B valuation with government backing [Funding]Britain bets on a homegrown inference chip to keep AI value at home.05Harvard, Anthropic and QuEra use AI to slash quantum computing costs [Research]AI-designed circuits cut the resources some quantum algorithms need by up to 39x.06NVIDIA, SK Telecom and Rebellions team up on Korean sovereign AI [Hardware]Korean-designed NPUs plug directly into NVIDIA's racks via NVLink Fusion.07Torus raises $500M for modular AI data centers that use the grid you already have [Funding]Prefabricated inference capacity online in weeks, not years, on existing power.08Meituan open-sources its trillion-parameter LongCat-2.5 under MIT license [Open Source]A vision-and-code model you can download, run privately and commercialize freely.09Australia to legislate mandatory safety rules for frontier AI companies [Policy]A new regulator and real penalties for any frontier lab that wants to operate there.10Berkeley and NVIDIA push 'agentic robotics' to program robots with AI teams [Robotics]Teams of AI agents write, test and fix robot code in simulation before deployment.
#01 · Research

OpenAI: frontier models are getting harder to monitor through their reasoning

# As models scale, the thoughts they show us explain less of what they actually do.
> the_news

OpenAI published research on November 10 studying how chain-of-thought (CoT) monitorability shifts across the model training lifecycle. Watching a model's step-by-step reasoning is one of OpenAI's main defenses against misbehavior, but the team found two worrying trends rising together: models can increasingly do hard tasks without reasoning out loud, and they get better at controlling their own CoT when prompted. In pretraining, lower base-model perplexity tracked with stronger no-CoT capability; in reinforcement learning on an Astra-class model, CoT controllability climbed sharply. The authors call the findings unsurprising but sobering, warning that future progress will need new approaches to keep reasoning legible.

> why_it_matters

Reading a model's chain of thought is how labs catch deception, reward hacking and misalignment before deployment. If that window closes as capability grows, safety teams lose their cheapest early-warning signal, and the whole industry needs new monitoring tools fast.

> real_life_use_case

An AI safety team that relies on reading an agent's reasoning trace to flag risky actions now knows that signal weakens as models get smarter, so it starts building activation-based monitors and adversarial evals instead of trusting the visible thought log alone.

#02 · Research

DeepMind's AlphaGenome 2 reads DNA cell by cell to pinpoint disease variants

# A next-generation model for predicting how specific cells interpret the genome.
> the_news

Google DeepMind unveiled AlphaGenome 2 on November 11, a major upgrade to its model for predicting how cells read DNA. The new version is markedly better at predicting gene regulation inside specific cell types, identifying nearly twice as many gene-expression-linked variants as the original AlphaGenome. By modelling the molecular effects of single-letter DNA changes at base-pair resolution across long sequences, it helps researchers separate the handful of genetic variants that actually drive disease from the vast background of harmless ones. DeepMind paired the release with resources like the AlphaGenome Atlas, a catalogue of molecular predictions for billions of possible single-nucleotide variants in the human genome.

> why_it_matters

Most disease-linked DNA changes sit outside genes, in hard-to-interpret regulatory regions. A model that predicts which variants matter in which cell types gives geneticists a shortcut from raw sequence to biological cause, speeding the hunt for the roots of complex diseases.

> real_life_use_case

A lab studying a heart condition feeds thousands of candidate variants into AlphaGenome 2, which flags the few predicted to disrupt gene activity specifically in cardiac cells, letting the team run wet-lab experiments on a short list instead of blindly testing everything.

#03 · Industry

Google launches the Gemini Decisions API for fast, calibrated AI judgments

# Score up to 1,000 yes/no, choice and rubric questions in a single multimodal pass.
> the_news

Google introduced the Gemini Decisions API on November 11, powered by a new purpose-built model, Gemini Decision 1. Instead of generating answers token by token, the model scores candidate options in a single forward pass, returning calibrated probabilities and confidence scores in 150-250 milliseconds. One request can mix text, images, audio, video and PDFs and ask up to 1,000 independent yes/no, multiple-choice or rubric questions. Priced at $0.10 per million input tokens with zero output tokens billed, Google says it is roughly 5x cheaper than using Gemini 3.5 Flash-Lite for the same work, and teams across Google already use it for model routing, content grading and policy checks.

> why_it_matters

A huge amount of real AI work is not writing prose but making bounded decisions: route this query, grade this answer, check this against policy. A fast, cheap, calibrated endpoint for exactly those calls lets builders replace brittle custom classifiers and cut both latency and cost.

> real_life_use_case

A support platform uses the Decisions API as a real-time router, classifying each incoming ticket in milliseconds to send simple questions to a cheap model and only the genuinely complex cases to a heavy reasoning model, trimming latency and spend across millions of requests.

#04 · Funding

UK AI chip startup Fractile hits $7B valuation with government backing

# Britain bets on a homegrown inference chip to keep AI value at home.
> the_news

Fractile, a London- and Bristol-based AI chip startup, was valued at $7 billion in its latest fundraising, Reuters reported on November 11, making it one of Britain's most valuable tech startups. The company raised a $720 million Series C led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, with backing from the British government. Fractile is developing chips designed specifically to speed up AI inference, the process of generating responses from trained models, while cutting energy use and cost. UK AI Minister Kanishka Narayan framed the investment as a bid to stop Britain inventing breakthrough technology only to see the industry and wealth built elsewhere.

> why_it_matters

Inference is where most AI compute spend now lands, and it is dominated by a few US suppliers. A well-funded challenger focused on cheaper, lower-power inference, backed by a national government, signals that countries increasingly treat AI silicon as strategic infrastructure, not just a commercial product.

> real_life_use_case

An enterprise running large models in production swaps some of its inference workload onto Fractile-class hardware designed for the job, aiming to serve the same traffic at lower energy cost and reduce its dependence on a single chip vendor.

#05 · Research

Harvard, Anthropic and QuEra use AI to slash quantum computing costs

# AI-designed circuits cut the resources some quantum algorithms need by up to 39x.
> the_news

QuEra Computing, Harvard and Anthropic highlighted research on November 10 showing that AI can design quantum circuits that need far fewer resources. Pairing Anthropic's Claude with an automated evaluator that scored each design on gate count and qubit count, the team improved seven of eight computational building blocks used in larger quantum algorithms. Results included a 12x cut in Toffoli gates for a key optimization step and a 39x reduction in a combined resource metric for a numerical function. Claude also discovered a new, more efficient approach to a core arithmetic operation. The work, led by Anthropic's STEM Fellows program, points to AI as a design partner for making near-term quantum hardware useful sooner.

> why_it_matters

Fault-tolerant quantum computers are bottlenecked by how many gates and qubits an algorithm demands. If AI can routinely redesign circuits to use a fraction of those resources, useful quantum applications arrive on smaller, nearer-term machines instead of waiting for hardware to catch up.

> real_life_use_case

A quantum chemistry team trying to simulate a molecule runs its circuit designs through an AI-plus-evaluator loop that returns a version using far fewer gates and qubits, letting the simulation fit on the modest fault-tolerant hardware actually available instead of a hypothetical future machine.

#06 · Hardware

NVIDIA, SK Telecom and Rebellions team up on Korean sovereign AI

# Korean-designed NPUs plug directly into NVIDIA's racks via NVLink Fusion.
> the_news

NVIDIA, SK Telecom and Korean AI chip startup Rebellions announced a collaboration on November 9 to build sovereign AI infrastructure in South Korea. Using NVIDIA's NVLink Fusion program, Rebellions will connect its next-generation R200 neural processing units to NVIDIA's open rack-scale infrastructure, with NVIDIA GPUs handling the prefill stage of inference and Rebellions NPUs handling decode to boost throughput. SK Telecom plans to deploy the combined systems in its AI data centers as part of a full-stack sovereign AI strategy spanning infrastructure, homegrown models and public services. South Korea's science ministry is backing the effort, with billions earmarked for domestic AI chips, data centers and a sovereign AI model.

> why_it_matters

Governments increasingly want AI built on their own soil, chips and models, not rented wholesale from abroad. Letting a national chip startup slot its NPUs into NVIDIA's dominant rack architecture is a template for how countries can pursue sovereign AI without starting their hardware stack from zero.

> real_life_use_case

A Korean public-service provider runs its citizen-facing AI on SK Telecom data centers where NVIDIA GPUs and local Rebellions NPUs share the inference workload, keeping sensitive data and the underlying hardware within national borders.

> source

NVIDIA ↗

#07 · Funding

Torus raises $500M for modular AI data centers that use the grid you already have

# Prefabricated inference capacity online in weeks, not years, on existing power.
> the_news

Salt Lake City-based Torus launched Foundry on November 11, a modular AI infrastructure product, alongside $500 million in growth capital led by Magnetar. Foundry packages power, storage, cooling, networking and compute into a prefabricated system delivering 1-20 megawatts per site in 12-16 weeks from contract to running workloads, versus the years a traditional hyperscale campus takes. The trick is using unused headroom on existing grid connections, paired with Torus's energy-storage product to shape loads, rather than building new generation or transmission. Torus says it has thousands of GPUs deployed today and plans more than 47,000 in Foundry units by the end of 2027, with a waterless cooling design that uses 95% less water than evaporative coolers.

> why_it_matters

Power, not chips, is now the main bottleneck for AI, and demand is shifting toward always-on inference close to users. A system that unlocks existing grid capacity and ships in weeks attacks the slowest, most expensive part of scaling AI compute.

> real_life_use_case

An AI cloud provider that can't wait years for a new data center installs Foundry units at an existing industrial site with spare grid capacity, bringing fresh inference compute online in a few months to serve a surge in customer demand.

#08 · Open Source

Meituan open-sources its trillion-parameter LongCat-2.5 under MIT license

# A vision-and-code model you can download, run privately and commercialize freely.
> the_news

Chinese delivery and e-commerce group Meituan released the open weights for LongCat-2.5 on November 9, about a week after introducing the multimodal model, publishing the checkpoint on Hugging Face under the permissive MIT license. The model activates about 48 billion parameters per token out of 1.6 trillion total and supports a one-million-token context, aimed at long-horizon agentic tasks like complex codebase development and visual coding. Meituan quantized expert weights to four bits for domestic AI accelerators, reporting over 40% higher throughput than an eight-bit baseline, and its hosted API lists a limited-time rate of $0.30 per million input and $1.20 per million output tokens. The MIT license permits commercial use, modification and redistribution.

> why_it_matters

An MIT-licensed, trillion-parameter vision-code model gives enterprises downloadable weights they can run inside their own infrastructure, automating visual testing and interface work without streaming proprietary code or UI recordings to an outside API. Permissive open weights at this scale keep pressure on closed frontier labs.

> real_life_use_case

A software team that can't send its private codebase to a third-party API downloads LongCat-2.5, runs it on internal hardware, and uses it to automate multi-step visual regression testing and UI prototyping entirely behind its own firewall.

#09 · Policy

Australia to legislate mandatory safety rules for frontier AI companies

# A new regulator and real penalties for any frontier lab that wants to operate there.
> the_news

The Australian government announced on November 10 that it will legislate mandatory safety and security protections for all frontier AI companies that want to operate in the country, enforced by a new targeted AI regulator working with the Australian AI Safety Institute. Under the proposed Australian AI Standards, companies would have to maintain security controls across development and deployment, identify and manage serious model risks, allow independent evaluation of their models, rapidly notify the government of incidents, and publish safety and governance frameworks. The regulator could impose significant penalties for non-compliance, and the rules are designed to give earlier warning of risks from rogue AI agents.

> why_it_matters

Most AI governance so far has been voluntary commitments. A wealthy democracy moving to binding, enforceable obligations, including mandatory third-party evaluation and incident reporting, sets a precedent other governments may copy and shapes how global frontier labs operate well beyond Australia.

> real_life_use_case

A frontier lab planning to offer its most capable model in Australia builds a compliance process: documented security controls, an independent evaluation before launch, and a rapid incident-notification pipeline, work it can reuse as similar rules spread to other markets.

#10 · Robotics

Berkeley and NVIDIA push 'agentic robotics' to program robots with AI teams

# Teams of AI agents write, test and fix robot code in simulation before deployment.
> the_news

UC Berkeley roboticist Ken Goldberg, with NVIDIA's Spencer Huang and colleagues, detailed a new approach they call 'agentic robotics' at the Conference on Robot Learning in early November. Instead of collecting massive datasets and hand-tuning every behavior, teams of AI agents generate robot-control programs, rehearse them across task variations in NVIDIA's Isaac simulation, identify where they fail and refine the code, with human engineers reviewing and verifying. The team released open-source code and an industry benchmark. In one case, giving OpenAI's Astra model a simulation interface and a short video let it build a physics simulation reproducing the clip, including realistic sponge deformation, in under an hour. Logistics firm Ambi Robotics applied the methods to lift per-robot sorting capacity by over 15,000 packages a year.

> why_it_matters

Two things have long slowed robotics: the enormous data needed to train robots and the heavy engineering to program them. Letting AI agents generate, test and repair robot code in simulation attacks both, and early real-world gains in warehouses suggest it is more than a lab demo.

> real_life_use_case

A warehouse operator uses agentic robotics to update how its sorting robots handle oddly shaped parcels, with AI agents generating and stress-testing the new behavior in simulation first, so the fleet gets a verified upgrade without weeks of hand-tuning on the floor.

# get this in your pocket every morning
Join on WhatsApp →
# AI Majlis Daily · Abu Dhabi · Dubai · Sharjah