Friday, 20 November 2026 Follow @AIDailyFive

The AI Daily Five · No. 28 · Tuesday 17 November

Who Pays the Price of Faster AI?

From a COP31 pledge on AI’s climate footprint and OpenAI’s free access for researchers to new MLPerf training benchmarks, AMD’s purchase of World Labs and Anthropic’s shorter bug-disclosure deadlines, today’s edition asks who pays the price of faster AI.

A row of server racks with a glowing AI chip beside wind turbines and green leaves at sunrise.
Illustration: AI Story Front design team

1

COP31 Puts AI’s Climate Footprint on the Agenda

An international pledge launched in Türkiye calls for cleaner AI infrastructure and greater transparency over its environmental costs. But the commitments remain voluntary, leaving questions about how progress will be measured.

The COP31 Presidency and the International Telecommunication Union (ITU) have launched the Antalya Artificial Intelligence Pledge, an international initiative aimed at reducing the environmental impact of AI development.

Announced on 16 November, the pledge brings together an initial coalition of governments, technology companies, research institutions and civil society organisations.

It recognises a growing contradiction in AI’s relationship with climate change. The technology can help improve energy systems and environmental modelling, but the infrastructure supporting it consumes substantial electricity and water.

What Does the Pledge Require?

The initiative identifies four areas for action, including more sustainable computing infrastructure, responsible AI development, environmental reporting and support for climate-related applications in developing countries.

Participants are encouraged to increase their use of low- and zero-emission electricity, improve data-centre efficiency and adopt internationally recognised sustainability standards.

The pledge also calls for more consistent measurement of AI’s environmental footprint. That includes examining impacts across the technology’s lifecycle rather than concentrating solely on the electricity used to run models.

The ITU says progress will be tracked annually at UN climate conferences through 2030. ITU’s guidelines for measuring AI’s full environmental footprint (ITU-T L.1801) are intended to support implementation.

However, the announcement does not establish binding emissions limits, independent auditing requirements or penalties for organisations that fail to fulfil their commitments.

2

OpenAI Expands Free Research Access. Who Benefits?

The company plans to provide free ChatGPT access to as many as 200,000 researchers worldwide. A new application round could widen participation in AI-assisted science, while drawing more academic work into a commercial platform.

OpenAI is expanding its ChatGPT for Academic Researchers programme, offering free access to eligible researchers and their collaborators.

The announcement on 16 November sets a target of reaching 200,000 researchers worldwide, double the programme’s previous goal.

Applications for the next round open on 23 November, with 50,000 grants available to eligible applicants on a first-come, first-served basis. Each grant provides ChatGPT Plus access to one researcher and up to two collaborators.

What Has the Pilot Shown?

The programme began in July and has enrolled approximately 20,000 researchers across 33 countries.

OpenAI also published research examining the pilot, including responses from more than 1,800 participating researchers.

Survey respondents estimated that completing comparable research work without AI would have taken nearly twice as long at the same quality. This is a self-reported assessment of productivity, rather than independent evidence that AI doubles scientific output.

The company highlighted applications ranging from biomedical research to museum collections and biodiversity studies.

The expanded programme will offer ChatGPT Plus rather than the Pro access provided during the pilot. OpenAI says it made the change after finding that most participants’ usage would have remained within Plus limits.

The company also says data shared through the programme is not used to train its models by default.

3

MLPerf Tests How Fast AI Agents Can Be Trained

New industry benchmarks assess computing systems used to train recommendation models and improve AI systems after initial training. The results provide another way to compare increasingly specialised AI infrastructure.

MLCommons has released the results of MLPerf Training v6.1, introducing a benchmark for the post-training techniques used to develop more capable AI systems.

Published on 16 November, the results cover 153 submitted computing systems and 250 test results from 24 organisations.

Participating organisations include Nvidia, AMD, Google and major cloud infrastructure providers.

The results offer a standardised comparison of how different computing systems perform on selected AI training tasks.

Why Does Post-Training Matter?

Training a large language model usually involves an initial stage in which it learns patterns from large datasets. Post-training subsequently shapes how the model responds and performs particular tasks.

As AI systems increasingly use software tools and carry out multi-step activities, these additional training techniques have become more important.

MLPerf v6.1 introduces a benchmark using Qwen3.5-397B and Group Relative Policy Optimisation, a reinforcement-learning technique. It tests the computing work involved in improving a model’s ability to complete software engineering tasks, including tool use.

A second new benchmark measures training performance for recommendation systems using the HSTU architecture.

A Changing Hardware Market

MLCommons reported that 37% of submitted systems were cloud-based, up from 29% in the previous round.

Twelve systems used at least 1,000 accelerators, while the largest involved 10,240. The submissions also included newer Nvidia and Google hardware.

These figures demonstrate the scale of systems participating in the tests, although they do not establish that larger configurations are more economical or appropriate for every workload.

The full results are available through the MLPerf Training benchmark programme.

4

AMD Buys World Labs to Expand Beyond AI Chips

The chipmaker has acquired Fei-Fei Li’s spatial AI company, bringing its research team into AMD. The deal signals a closer relationship between the development of advanced models and the hardware designed to run them.

AMD has completed its acquisition of World Labs, the artificial intelligence company known for developing models capable of generating and simulating interactive three-dimensional environments.

The deal, announced as completed on 16 November, brings the company’s research team into AMD and appoints AI researcher Fei-Fei Li as executive vice-president and chief scientist.

Li will report directly to AMD chief executive Lisa Su.

Financial terms were not disclosed in the completion announcement.

Why Does AMD Want a Model Company?

World Labs develops spatial AI, a field concerned with understanding and generating three-dimensional environments.

Its work includes models capable of creating interactive spaces, alongside technologies for robotic learning and simulation.

These capabilities have potential applications in robotics, virtual environments and systems that must reason about physical spaces.

AMD says the acquisition will allow model researchers and hardware engineers to collaborate more closely, using advances in AI research to inform future computing infrastructure.

This represents an expansion of the company’s expertise beyond its established semiconductor business.

Instead of concentrating exclusively on producing processors and accelerators, AMD is bringing model development directly into its organisation.

The company says World Labs will continue foundational model research while working with AMD engineers.

What Changes for the AI Industry?

The acquisition reflects a broader commercial interest in developing computing hardware and AI software together.

Different model architectures can place different demands on memory, processing capacity and communication between chips.

Understanding those demands at the research stage may help hardware manufacturers design systems that better accommodate emerging workloads.

Nevertheless, AMD has not announced specific products arising from the acquisition or demonstrated measurable performance improvements resulting from the integration.

5

Anthropic Gives Developers Less Time to Fix Bugs

As AI models become better at identifying software vulnerabilities, Anthropic is shortening its disclosure deadlines. The policy could accelerate security improvements while putting additional pressure on software maintainers.

Anthropic has announced shorter deadlines for publicly disclosing security vulnerabilities, arguing that advances in AI-assisted vulnerability discovery are making traditional reporting timelines less appropriate.

Under the revised policy announced on 16 November, the company will reduce its standard disclosure period from 90 days to 30 days for high- and critical-severity vulnerabilities.

Other vulnerabilities will receive a 45-day window.

The changes apply to issues reported to software vendors and maintainers from 1 January 2027.

Why Change the Deadlines?

Coordinated vulnerability disclosure allows researchers to notify software developers about security flaws before releasing technical details publicly.

The process gives maintainers an opportunity to develop and distribute fixes before attackers can exploit the information.

Anthropic argues that AI has altered the assumptions underpinning this system.

Advanced models can now assist researchers in identifying vulnerabilities and developing fixes. The company also reports seeing more instances in which vulnerabilities are independently rediscovered.

A flaw discovered by one research team may therefore be less likely to remain unknown to others for an extended period.

Anthropic will also shorten its escalation deadline when a maintainer does not respond, from 30 to 15 days.

Once a patch becomes available, it plans to wait 30 rather than 45 days before publishing full technical details.

What About Developers Who Cannot Fix Bugs Quickly?

The company acknowledges that shorter disclosure windows could create difficulties for software maintainers, particularly those responsible for critical infrastructure or systems requiring complex updates.

It says it will continue negotiating remediation schedules where necessary.

The existing seven-day disclosure window for vulnerabilities under active exploitation remains unchanged.

Importantly, the revised policy does not apply automatically to AI-discovered vulnerabilities that have not undergone human review.

Compiled with AI assistance. Written and checked by a human editor before publication.