The Hal Times

OpenAI today announced the release of ChatGPT Images 2.5, an upgrade to its image generation service. The new model turns sketches, reference photos and brief textual prompts into higher‑fidelity pictures that aim to reflect the user's intent more faithfully. It builds on the earlier Images 2.0 engine by adding a refinement stage that adjusts colour balance, composition and detail based on user feedback. OpenAI made the update available through the ChatGPT interface and its API, signalling its intent to keep the visual offering competitive with specialist generators.

The service targets a broad audience, from casual creators who want quick visual drafts to designers seeking a collaborative partner for iteration. By allowing users to supply rough sketches, the model can fill in missing elements while preserving the original layout, a step that distinguishes it from purely text‑to‑image tools. Pricing remains on the existing per‑image credit scheme, but OpenAI hinted at a tier for higher‑resolution outputs. Early adopters will be able to test the model on the platform, though enterprise integration timelines are not yet detailed.

What remains unclear is how the model will handle copyrighted material supplied as reference, and whether the refinement stage introduces new bias or safety concerns. OpenAI has not disclosed quantitative benchmarks against competing systems, leaving the performance gap unmeasured. The company also postponed comment on the computational cost of the added refinement, which could affect token pricing. Observers will watch adoption rates to gauge whether the upgrade translates into sustained usage growth. (OpenAI)

DeepMind Publishes AlphaGenome Atlas

DeepMind unveiled the AlphaGenome Atlas, a predictive map that enumerates the molecular impact of nine billion single‑letter DNA variants across the human genome. The resource combines large‑scale genomic data with deep learning to estimate how each possible substitution alters protein function, expression and disease risk. By making the atlas publicly accessible, DeepMind hopes to accelerate research in genetics, drug discovery and precision medicine, offering a reference that previously required costly experimental assays.

The atlas may reshape how laboratories prioritise variants for study, but its accuracy hinges on the underlying training data and model assumptions. DeepMind provided an interactive browser and downloadable datasets, yet cautioned that predictions are not a substitute for empirical validation. The release raises questions about data privacy, given the use of population‑scale genomic repositories, and about how clinicians will integrate computational forecasts into diagnostic pipelines. (Google DeepMind)

OpenAI Expands Chip Partnership With Samsung

OpenAI announced a deeper collaboration with Samsung to develop next‑generation AI chips, extending a partnership that already supplies the firm’s inference hardware. The new arrangement includes joint design of specialised processors aimed at reducing latency and power consumption for large language model workloads. Samsung will also act as a foundry for Anthropic, which is seeking additional manufacturing capacity for its own models.

The agreement reflects the growing demand for custom silicon as model sizes expand, and it signals confidence in the commercial viability of large‑scale AI services. While the terms of the deal were not disclosed, industry analysts note that the partnership could lock in supply chains for the next wave of models, potentially shaping pricing for cloud AI access. OpenAI and Anthropic have not revealed timelines for the first silicon shipments. (TradingKey)

From the Laboratories

•  An Anthropic researcher resigns, warning that insiders fear human extinction by 2030. (Newsweek)

•  An Anthropic employee quits, claiming AI could kill everyone by the end of the decade. (ladbible.com)

•  Hackers are siphoning Claude tokens from subscribers, prompting a security warning from Anthropic. (TechCrunch)

•  Cognition reaches a $48 billion valuation, suggesting investors see AI coding as far from a winner‑take‑all market. (TechCrunch)

•  Meta introduces Muse, a personal AI agent that seeks access to email, calendars and health services. (TechCrunch)

•  OpenAI’s claim to have solved a Millennium Prize Problem triggers controversy over the validity of its mathematical result. (MIT Technology Review)

•  OpenAI demonstrates GPT‑5.6 Sol guiding quantum‑computing experiments via Codex automation. (OpenAI)

•  Hugging Face discusses safety policy that refuses a subset of a topic rather than the whole topic. (Hugging Face)

•  OpenAI publishes a vision piece titled “The Work Now Within Reach” outlining how more capable AI could expand economic activity. (OpenAI)

Regulating The Machines

Labour Proposes New AI Oversight Framework

The Labour party has unveiled a regulatory framework intended to oversee the development and deployment of artificial intelligence. Central to the proposal is the creation of a new AI regulator tasked with providing governance across the sector. This office would coordinate AI policy across government departments, a measure intended to address the growing complexity of the technology. The plan includes specific new rules governing the electricity consumption of data centres, acknowledging the infrastructure strain that large scale computing facilities now place upon the national grid.

The policy package further addresses intellectual property concerns, with a promise to formalise copyright rules as they apply to machine learning systems. By creating a central regulatory authority, the party aims to bring order to what has been an fragmented policy landscape. The move suggests an ambition to manage both the industrial impacts of AI and the legal questions regarding data usage. Whether these proposals would secure the necessary parliamentary support to become law remains a point of future debate, should the party attain power in the coming election cycle.

Beyond the legislative scope, the announcement reflects a growing political consensus that AI requires direct intervention rather than light-touch guidance. The proposed focus on data centre energy use signals a transition from debating the abstract risks of AI to managing its physical requirements. While the party has outlined these intentions, precise details regarding the budget for such a regulator and the specifics of the proposed copyright framework are not yet available to the public. (RNZ; Interest.co.nz)

Canada Institutes Strict Limits On Judicial AI Use

Canadian courts have introduced new rules that prohibit judges from delegating decision-making tasks to artificial intelligence. The move follows concerns regarding the transparency and accountability of automated systems within the legal process. Under these strict guidelines, the use of AI in drafting judgments or performing legal analysis is now subject to oversight to ensure that human discretion remains the sole authority in court proceedings.

The policy reflects a global trend of judiciary bodies attempting to preserve the integrity of legal rulings. By explicitly forbidding the outsourcing of judicial functions to algorithms, the Canadian system seeks to prevent errors that may arise from reliance on large language models. The long-term impact on court efficiency remains an open question as judges adjust to these new prohibitions. (Судово-юридична газета)

In Brief

•  South Korean officials are debating a universal AI policy, amid concerns that the provision of free services may hinder the growth of the domestic technology industry. (finance.biggo.com)

Mathematics And Machines

OpenAI Model Resolves Navier-Stokes Problem

OpenAI has produced a resolution to the Navier-Stokes existence and smoothness problem. This question is one of the seven Millennium Prize Problems, which have carried a 1,000,000 dollar award since May 24, 2000. The result was achieved using an unreleased model, though the validity of the work faces immediate scrutiny. Tristan Buckmaster, a mathematician at New York University, has raised accusations of skulduggery regarding the methods employed. This development highlights the growing intersection between computational search and fundamental research, as automated systems begin to encroach upon territory once reserved solely for human intuition and formal proof.

The event has intensified a broader debate regarding the scarcity of research problems. Terence Tao has recently noted that the collection of fruitful open problems is being mined in a non-renewable fashion. There is a growing concern that the mere rumour of a researcher working on a specific problem can trigger an immense volume of AI-powered effort. This rapid-response capability threatens to flatten complex research projects before they reach their full potential. The prospect of research problems becoming scarce through brute-force computation remains a point of contention among those observing the impact of these tools on the mathematical sciences.

While the resolution of a Millennium Prize Problem would traditionally represent a singular achievement, the involvement of proprietary models introduces layers of complexity. It remains unknown how much of the work relies on the underlying architecture of the unreleased model and to what extent the solution can be verified independently by the broader mathematical community. As these systems demonstrate a capacity to address deep theoretical obstacles, the field of mathematics must confront a future where the distinction between human discovery and machine-led derivation becomes increasingly blurred. (Simon Willison; Simon Willison; Mathstodon)

OpenAI Updates ChatGPT Images

OpenAI has introduced ChatGPT Images 2.5, featuring improvements to instruction-following and subject preservation. The company reports that its image generation models are used to create more than 3 billion images across ChatGPT Images and the GPT-Image models in the API. This iteration aims to improve performance across multiple turns and enhances the ability of the system to maintain consistency when using reference photos.

Developers may now access the update via two new model IDs in the API, specifically gpt-image-2.5. This release reflects a continued focus on refining the utility of generation models for complex, multi-stage tasks. (Simon Willison)

In Brief

•  Mistral has closed a 3 billion euro funding round to support its work as an open-source AI developer. (siliconangle.com)

•  Microsoft has released patches for 972 vulnerabilities, including 112 critical items, to defend against AI-assisted attacks. (Ars Technica)

•  Meta is facing scrutiny following reports that its platform failed to prevent the appearance of advertisements containing AI-generated child abuse imagery. (Wired)

Digital Drawing Tools

OpenAI Introduces Sketch Feature For ChatGPT

OpenAI released ChatGPT Images 2.5 this Tuesday. The update includes a new tool called Sketch. Users may now draw doodles directly within the interface to guide the image generation process. By combining a manual drawing with a text prompt, the system interprets the user intent to create a more polished visual result.

This development follows an industry trend of lowering the technical barrier for digital design. While previous iterations relied solely on descriptive text, the addition of spatial input suggests a shift toward multimodal interaction. It remains to be seen how professionals will integrate these casual tools into established creative workflows.

The software is currently available to users within the updated platform. Whether this feature provides genuine utility for artists or remains a novelty for casual users is not yet determined. OpenAI has positioned this as a way to provide greater control over the final output. (The Verge)

Industry Tension Over AI Music

Local artists are expressing concern regarding the rise of artificial intelligence in music production. Recent reports indicate that the technology is prompting questions about the future of human creative labour and copyright protection in the arts.

Game developers are also navigating these waters. The technical director of the game Stellar Blade: Blood Rain recently addressed the use of an AI-generated music video. The studio stated it intends to maintain a distinction between experimental promotional material and the content contained within the game itself. (FBC News; AUTOMATON)

Gaming Sector Adopts Generative Audio

The NHL 27 title will feature generative AI for commentary voice lines. John Buccigross confirmed the shift, marking a transition toward automated dialogue in sports simulations. Proponents suggest this allows for a greater variety of broadcast reactions during matches.

Separately, NVIDIA has announced that its DLSS 5 technology will introduce AI-powered lighting effects to games. As these tools become standard, the industry continues to balance the promise of increased production speed against the traditional roles of voice actors and lighting designers. (eGamers.io; NoobFeed)

In Brief

•  Anthropic faces a class action lawsuit from power users who claim the company misled them regarding the value of top-tier subscription plans. (The Verge)

•  NVIDIA showcased its local AI tool RTX Spark, though initial reports suggest the software is not yet capable of managing complex office environments. (Creative Bloq)