OpenAI launched Dots, its new always-on agents

Good day, AI enthusiasts!

This week’s AI news was unusually broad: a major acquisition, another huge data-centre project, record demand for memory chips and several moves showing where AI is becoming more specialised.

Here are 10 AI developments worth knowing from the last seven days:

AMD announced on 28 September that it has agreed to acquire World Labs, the spatial-intelligence research company founded by AI researcher Fei-Fei Li, in an all-stock deal worth approximately $8.2 billion.

World Labs is developing AI systems capable of understanding and generating three-dimensional environments. Li is expected to join AMD as executive vice president and chief scientist once the deal closes, which AMD expects by the end of 2026.

What makes the acquisition interesting is where AMD is spending the money. Rather than buying another semiconductor company, it is bringing frontier model research directly into a chipmaker — giving its hardware, software and research teams a much closer view of the workloads future AI systems may require.

Source: AMD newsroom

Japan’s largest power generator, JERA, signed an agreement with Dell Technologies and RHAELM on 1 October to develop a standardised model for large-scale AI infrastructure in Japan.

The first project, at JERA’s Chiba Thermal Power Station, is designed for up to 400MW of capacity and is expected to require more than $15 billion in total capital deployment. Operations are targeted to begin around 2028.

The location is particularly significant. The project will sit alongside an existing power-generation asset rather than depending entirely on conventional grid connections. As AI data centres grow, access to power — and how quickly it can be delivered — is increasingly influencing where infrastructure gets built.

Source: JERA

Micron reported record fiscal fourth-quarter revenue of $54.23 billion on 30 September, compared with $11.32 billion in the same quarter a year earlier.

The company explicitly pointed to AI-driven demand as a major driver and said it enters fiscal 2027 expecting another record year.

The scale of the increase is a useful reminder that the AI infrastructure boom reaches well beyond GPUs. High-bandwidth memory and other memory products are essential to running increasingly large AI systems, making suppliers such as Micron part of the same investment cycle.

Source: Micron investor relations

Google introduced Gemini 4 Argon on 30 September, positioning the frontier model around long-running software engineering, professional knowledge work and cybersecurity.

One of the more striking technical changes is its one-million-token output limit, up from 64,000 tokens previously. Google says that allows the model to sustain much longer reasoning and generation trajectories.

But the rollout is deliberately limited. Argon is initially going to selected cybersecurity defenders through Google’s Fairwind programme while the company continues testing before broader developer, enterprise and consumer access.

Introductory API pricing is set at $2 per million input tokens and $10 per million output tokens.

Source: Google

OpenAI introduced Dots at DevDay on 29 September: persistent agents with connected apps and their own cloud computers.

Unlike a normal assistant session, a Dot can be assigned continuing work rather than starting again from scratch every time someone opens ChatGPT. OpenAI is also previewing “specialist dots” for organisations, where agents are assigned particular responsibilities.

This is a more meaningful distinction than simply making chatbots smarter. Persistent agents introduce practical questions around permissions, supervision, connected systems and ownership of ongoing work — precisely the issues businesses have to solve if agents are going to move beyond isolated tasks.

Source: OpenAI Developer Community

Synopsys announced a partnership with OpenAI on 30 September to develop GPT-Synopsys, a specialised model for semiconductor design.

The project combines OpenAI’s frontier models with Synopsys’ electronic-design automation tools and chip-engineering expertise.

This is a good example of where specialist AI may become commercially important. Chip design involves highly technical workflows where general reasoning alone is not enough; models also need domain-specific engineering context and access to tools capable of checking whether their output actually works.

For Synopsys, it also creates another way to monetise AI alongside its existing engineering software and agent products.

Source: Synopsys investor relations

Shopify extended its WebMCP support on 28 September so compatible browser-based AI agents can now interact directly with an active checkout.

Agents can read checkout information, update supported fields and ultimately submit the checkout using Shopify’s complete_checkout tool — but only after buyer confirmation. When authentication or other direct user input is required, control is handed back to the shopper.

It takes agentic commerce a meaningful step beyond product recommendations. An AI can now help a shopper move from discovery through cart management and checkout inside the same browser workflow.

The emphasis on confirmation is equally important: Shopify is expanding what agents can do without removing the buyer from the final transaction.

Source: Shopify Developers

The open-source TileLang project added an official Huawei Ascend 950 backend on 30 September.

The update includes native code generation, automatic scheduling and synchronisation, SIMD/SIMT vector programming and examples for workloads including matrix multiplication and FlashAttention.

Hardware competition in AI is partly a software problem. Nvidia’s strength comes not just from its chips but from the mature developer ecosystem surrounding them. Better tooling for alternative accelerators makes those chips considerably easier for researchers and companies to use.

That makes projects such as TileLang an important — if less visible — part of the wider competition around AI compute.

Source: TileLang on GitHub

Anthropic published new research on 30 September examining which physical work tasks present-day robots can perform.

Its researchers estimate that robots can perform physical tasks representing around 34% of total US working time in at least some circumstances. But when economics are included, robots are currently cost-competitive with human labour for just 0.3% of job tasks.

That distinction matters. Technical capability is not the same as adoption, and neither automatically translates into job displacement.

Anthropic also identified physical capability, regulation and human preferences as additional barriers. The study is therefore less a prediction of imminent automation than an attempt to measure where robotics actually stands today.

Source: Anthropic Research

California’s Department of Justice announced on 1 October that Attorney General Rob Bonta had served OpenAI with an investigative subpoena.

The subpoena forms part of an existing investigation into cybersecurity incidents and risks involving OpenAI and its models.

An investigative subpoena is not a finding that OpenAI broke the law. It gives the state additional powers to obtain information while it examines what happened and whether existing California laws apply.

It is another indication that regulators are beginning to examine not just AI outputs but how frontier models behave when connected to tools, systems and cybersecurity workflows.

Source: California Department of Justice

So what did this week actually tell us?

Infrastructure remained enormous. AMD committed $8.2 billion to bring World Labs inside the company, Japan moved forward with a $15 billion-plus AI infrastructure project and Micron’s results showed what the compute boom is doing to memory demand.

At the same time, models are becoming more specialised. Google is positioning Argon around difficult professional and cybersecurity work, while OpenAI and Synopsys are developing a model specifically for semiconductor design.

Agents also moved closer to real commercial activity. OpenAI’s Dots are designed to remain active over time, while Shopify is giving browser agents a controlled route through checkout.

O T H E R__A I _N E W S

Cloudflare released Clef and Clef-flash, two open-source “decision models” designed to make fast, bounded choices rather than generate long-form responses. Cloudflare is also introducing reinforcement-learning fine-tuning for this type of workload.

Strands Agents released Strands Decider 2B, a two-billion-parameter open-source decision model designed for local development and tasks such as routing, classification and tool selection.

Reco raised another $55 million, bringing total funding for the AI-agent security company to $140 million. The new round includes participation from AT&T Ventures.

Utopai Studios introduced its new PAI platform and Utopai X video model for film and television production. Utopai says X ranked second globally on Artificial Analysis’ text-to-video leaderboard with audio in the 29 September results.

Anthropic published “Claude-shaped science”, a guest research piece from Harvard physicist Matthew Schwartz exploring how AI can accelerate certain forms of quantitative scientific work and describing the BootLoops calculation toolkit.

Cloudflare opened the waitlist for managed Cloudflare OS deployments. The company describes OS as an organisational agent workspace that can work with company context, data and systems; the managed deployment is the new development this week rather than the original OS announcement.

ServiceNow CoreAI published AutoSynthData, a system that turns weaknesses found in enterprise AI agents into new validated training tasks. In one reported experiment, synthetic fine-tuning improved mean Pass@1 by 7.2 percentage points.

OpenAI introduced reusable Codex cloud environments, allowing coding agents to continue working after a laptop is closed and letting users monitor or steer tasks from another device.

See you soon,

News Digest AI team