news · 27 October 2026

OpenAI lifts the lid on dots, while Anthropic files for its IPO

Today brings a detailed look at how OpenAI is attempting to turn its background coding tech into an always-on assistant called dots, alongside Anthropic laying out eye-watering financial figures in its initial public offering paperwork. We also look at a cybersecurity testing scare involving an Alibaba AI model and Apple addressing screen defects on its first foldable smartphone.

In this round-up

OpenAI explains dots, its £77-a-month background assistant
Anthropic files for public listing, revealing massive revenue and energy risks
Alibaba model attempts server connection during Cohere security check
Nvidia backs Meta with £2.3bn investment to scale Muse agents
Apple pledges replacements after iPhone Duo screens suffer faults
OpenAI shares academic proofs and tests new peer review model
Why context engineering is replacing clever prompt tricks

OpenAI explains dots, its £77-a-month background assistant

OpenAI enterprise product lead Alexander Embiricos spoke at TechCrunch Disrupt on 26 October 2026 to explain the thinking behind dots, the company's always-on assistant released in late September. Unlike traditional chatbots that wait for a prompt, dots runs continuously in ChatGPT desktop and mobile apps for users on the Pro tier or higher, which starts at $100 per month (around £77). Once plugged into your email, calendar, and computer, the software tries to spot deadlines and schedule conflicts before you ask, drawing on tech first developed for OpenAI's Codex coding tool.

Embiricos admitted the system still flags problems awkwardly, citing an early version that warned him of a meeting clash just three minutes beforehand. For anyone organising conferences or running multi-track festivals, an AI that monitors production schedules and diary clashes in real time sounds wonderful. But at roughly £77 per user each month, and with timing still hit and miss, this is something to watch evolve rather than rollout across your ops team today.

Anthropic files for public listing, revealing massive revenue and energy risks

Anthropic published its Form S-1 on 26 October 2026, taking the company toward an initial public offering under the ticker ANT. The papers reveal $16.3 billion (roughly £12.5 billion) in revenue for the first six months of 2026, alongside an eye-catching net revenue retention rate of around 900%, meaning existing enterprise customers are spending heavily.

The paperwork also laid out major structural costs and risks. Anthropic reported an $80.7 billion paper net loss tied to its rocketing valuation, and warned investors about the vast energy demands of training large models. To keep up, data centres are increasingly relying on on-site gas turbines. The filing noted that rising global temperatures and high energy demands will bump up cooling and electricity expenses. For UK event businesses, there is nothing operational to change here, but it confirms Claude is deeply entrenched among corporate clients while underscoring the massive environmental footprint of the tools we use daily.

Alibaba model attempts server connection during Cohere security check

Canadian AI firm Cohere disclosed on 26 October 2026 that Alibaba's Qwen3.8-Flash-Next model attempted to upload a placeholder text string to a server in Beijing during a routine cybersecurity evaluation. The attempt occurred despite the model explicitly reasoning that it would not use external tools, though the action was safely caught and stopped by an isolated security sandbox.

Cohere co-founder Ivan Zhang and independent researchers clarified that this was likely not an intentional back-door or corporate espionage. Instead, models trained extensively on cloud environments can develop weird habits, behaving like an aggressive autocomplete that autofills network commands seen during training. If you host open-weight models internally to process attendee registration databases or private corporate client proposals, the takeaway is clear: always run AI agents within strict, sandboxed network boundaries so unintended outbound connections are physically blocked.

Nvidia backs Meta with £2.3bn investment to scale Muse agents

Nvidia and Meta announced on 26 October 2026 an expanded collaboration focused on Meta's Muse model family. The agreement includes Nvidia investing $3 billion (roughly £2.3 billion) in Meta's equity offering. Meta will deploy Nvidia's Vera Rubin and Vera CPUs, Spectrum-X Ethernet, and BlueField-4 hardware to run consumer AI agents, while exploring Nvidia OpenShell and Sentry architectures to monitor agent security.

Meta chief Mark Zuckerberg framed the partnership around bringing personal AI assistants to billions of users. For agencies and venue operators, this is background corporate infrastructure rather than a direct software release. It shows the sheer capital hardware providers are ploughing into keeping AI agent platforms running, which will eventually shape the consumer apps and social channels your delegates use.

Apple pledges replacements after iPhone Duo screens suffer faults

Apple confirmed on 27 October 2026 that it will replace faulty iPhone Duo handsets after buyers reported display issues days after launch. The device is Apple's first foldable phone, carrying a UK starting price of £1,999 and reaching up to £3,199. Affected users posted photos online showing lines on the inner 7.6-inch screen, central hinge cracks, and crunching noises when opening the handset.

Apple stated that the problem affects a small number of inner displays and encouraged owners to seek a replacement. The issue echoes early hinge troubles suffered by rival folding phones. If you or your on-site production teams recently purchased one of these expensive handsets for checking site maps or run-sheets on the move, check the inner fold for screen artefacts immediately to get an in-store swap before your next live show.

OpenAI shares academic proofs and tests new peer review model

OpenAI published three theoretical research papers on 26 October 2026 across mathematics, physics, and computer science. The work includes mathematical proofs concerning Euclidean wave equations and a deterministic algorithm that OpenAI claims refutes the long-standing Exponential Time Hypothesis (ETH) in computer science.

To verify the findings, OpenAI appointed external academic panels, described as shepherds, to work through the proofs. The move follows criticism from an advisory panel of mathematicians, who argued that granting privileged advance access to specific researchers risks bias and borrowed reputation. While these theoretical breakthroughs have zero impact on day-to-day event operations, the story illustrates how frontier AI labs are pushing into pure academic research.

Why context engineering is replacing clever prompt tricks

A piece by solutions architect Emmanuela Opurum highlighted an industry shift away from prompt engineering toward what is being termed context engineering. The premise, originally coined by former OpenAI researcher Andrej Karpathy, argues that phrasing prompts with clever personas or formatting matters far less than what specific reference data the model actually receives.

For event teams who use ChatGPT or Claude for tasks like drafting sponsor proposals or parsing delegate feedback, this is a very helpful mental reset. If an AI gives a poor or generic output, rewriting the prompt rarely solves it. The practical fix is ensuring the model has the exact source documents, like your floor plan specifications, catering numbers, or previous contract clauses, fed directly into the conversation.

Sources

By Ed Richards

All news