AI Daily Brief · Oct 18, 2026 | A fruit-fly brain flies through an AI-generated world; Microsoft reportedly quits Nscale's UK data center; mystery aurora models surface in OpenAI's API

It was a quiet Sunday in AI, and one of the most-engaged posts from an AI company over the past 24 hours came from Odyssey, which showed a fruit-fly brain model flying on its own through scenes generated by its Odyssey-3 world model. The Telegraph reported that Microsoft has walked away from Nscale’s AI data center project in Essex, England, because the grid can’t deliver power until the 2030s. Three aurora models showed up in OpenAI’s API model list and were then pulled, and for now they look more like image models. DeepSeek V4.1-Flash-1017 got its first outside tests, and Google Research open-sourced Harness², which lets agents improve their own harness. The gaming world’s AI fight also rolled on: one studio canceled a public demo for fear of AI clones, while Epic’s CEO predicted a new golden age of games.

Today’s highlights

• Odyssey placed a digital model of a fruit-fly connectome inside its Odyssey-3 world model and let it fly through an AI-generated harbor scene; the company’s post drew 11.7K likes and 2.41M views.
• The Telegraph reported exclusively that Microsoft ended its agreement for Nscale’s Loughton data center in Essex because grid power won’t arrive until the 2030s. Nscale says it is pressing ahead; Microsoft did not comment.
• Three models, aurora-red/green/blue-alpha, appeared in OpenAI’s API and then vanished. After one turned up in a model arena, they look like image models. OpenAI has said nothing.
• In Browser Use’s in-house tests, DeepSeek V4.1-Flash-1017 matched Opus 5.5 on long browser tasks at about a tenth of the cost per task.
• Google Research open-sourced Harness², which has agents propose, test and combine improvements to their own harness without touching model weights.

1. Odyssey puts a fruit-fly brain model inside its Odyssey-3 world model and lets it fly through AI-generated scenes

At 06:44 Beijing time on Oct 18, AI lab Odyssey posted that it had put a “fruit fly brain” into its Odyssey-3 world model, where the fly flies around on its own in a simulated world imagined entirely by the model. The post drew 11.7K likes and 2.41M views, making it one of the most-engaged posts from an AI company in the past 24 hours. Co-founder Oliver Cameron reposted it, telling the little fly to fly free.

According to Inshorts, Odyssey used a digital model of the fruit-fly connectome (a map of its neural connections) to control flight in the virtual environment; the demo shows a harbor with boats, buildings and pedestrians. Odyssey-3 is the world model Odyssey released on Oct 8, which the company describes as an autoregressive diffusion Transformer that generates interactive environments in real time in response to actions. Odyssey has not published technical details of the demo, such as which connectome it used or how the neurons are simulated.

Why it matters: A commentary from cAImpare cautions that a connectome is a map, not a mind, and that running one in software is only an engineering approximation of neural activity. The demo is better read as a proof of concept.

Reactions: @cgtwts said this is basically the simulation hypothesis (27.3K likes, 1.08M views). @theo joked that no fly can ever again be sure whether it lives in the real world or a simulation, which must be depressing (4,272 likes) [Joke]. Ethan Mollick called it both fun and instinctively unsettling (1,076 likes).

Sources: @odysseyml · @olivercameron · @cgtwts · @theo · @emollick · @kimmonismus · Odyssey-3 launch blog · Inshorts · cAImpare

2. Microsoft pulls out of Nscale’s AI data center in Essex, England, because the grid can’t supply power until the 2030s

The report [Media report; Microsoft did not comment]: In an exclusive published on Oct 18 (13:30 Beijing time), the Telegraph reported that Microsoft has told UK neocloud company Nscale it is ending their agreement for a data center in Loughton, Essex. Microsoft was the project’s anchor tenant and its only public customer, with a deal to rent about 23,000 Nvidia AI chips. The site was meant to come online in 2027, but the grid operator told Nscale power would not be available until the 2030s. Bloomberg, Reuters and the Guardian followed up.

What each side says: An Nscale spokesperson said the company is still pushing ahead with Loughton and is looking for alternative power, adding that its relationship with Microsoft remains strong and long-term and that this London-area grid issue does not affect other contracts. According to the Guardian, Microsoft is still involved in Nscale’s project in Norway. Microsoft did not respond.

Why it matters: Nscale filed for a US IPO last month, targeting a valuation of up to about $35 billion. According to the Telegraph, more than 40% of Nscale’s $103 billion in signed contract revenue comes from Microsoft, though the UK project accounts for only a small part of it; Anthropic is also a major customer. The UK government once called the project “the UK’s largest sovereign AI data center,” and it was part of Microsoft’s £22 billion UK investment pledge last year. In April, OpenAI also paused its Stargate project with Nscale in Tyneside, citing high electricity prices.

Reactions: UK user @s8mb lamented a project losing billions of pounds simply because it couldn’t get a grid connection for five years (2,686 likes). @BlackPantherCap argued that the market is mistaking CoreWeave’s financing problems for a problem with AI demand (personal opinion).

Sources: @jamestitcomb (Telegraph reporter) · @business · @ftr_investors · @s8mb · @BlackPantherCap · Telegraph via Yahoo · The Guardian · CNA (Reuters) · Bloomberg · Euronews (September power delay)

3. Three aurora models appear and vanish in OpenAI’s API, and look more like image models; talk of “Fable 5.5 next week” spreads too

On the morning of Oct 18, Beijing time, @scaling01 and then @AndrewCurran_ spotted three new entries in OpenAI’s API model list: aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha (AndrewCurran_’s post drew 2,005 likes). At 16:49, @scaling01 reported they were gone again, and @testingcatalog said the entries were pulled quickly and looked like an accidental release. At 17:26, LuminaBench said it had drawn aurora-red-alpha in a model arena and that it appeared to be an image model; testingcatalog then revised its view, saying they may be new image models.

Speculation [Unverified]: Earlier guesses ranged from a dedicated Codex coding model to a high-speed reasoning variant. @kimmonismus guessed a new Astra model aimed at Anthropic, and @ChrisGPT said these are OpenAI’s answer to Fable 5.5. OpenAI has published no documentation, pricing or explanation, and Crypto Briefing noted that OpenAI did not comment.

The Fable 5.5 rumor [Unverified]: @imjustnewatai claimed it had just been confirmed that Fable 5.5 arrives next week and that OpenAI has a response ready (853 likes), without citing a source. Anthropic has made no official announcement.

Also: Citing Similarweb data, @testingcatalog said OpenAI’s Tibo (@thsottiaux) had the most popular profile page on X in September by traffic share, ahead of Fabrizio Romano and Elon Musk (third-party data, 409 likes). He credited his “usage limit reset” posts.

Sources: @scaling01 1 · @AndrewCurran_ · @scaling01 2 (pulled) · @testingcatalog 1 · @LuminaBench · @testingcatalog 2 (image model) · @kimmonismus · @ChrisGPT · @imjustnewatai · @testingcatalog 3 (Similarweb) · Crypto Briefing

4. First outside tests of DeepSeek V4.1-Flash-1017: on par with Opus 5.5 on browser tasks at about a tenth of the cost per task

As of this issue’s compilation, the third-party leaderboard Artificial Analysis still lists only the September version of V4.1 Flash; the 1017 release has not been ranked yet. These are the outside tests available so far.

Browser Use: In in-house testing by the browser-agent company, the 1017 release scored 75/100 on long browser workflows, the same as Opus 5.5 and close to Astra. It cost $0.22 per task versus $2.35 for Opus 5.5 (about a 10x gap). Browser Use said Sol is still first, but not by much (414 likes).

Paweł Huryn’s personal benchmark [Personal test]: Using DeepSeek’s own harness at the max setting, the model averaged 42.7 over seven runs; with the Claude Code harness it scored 39.1. For comparison: GPT-6 Astra 45, Muse Wonder Full 1.0 42, Opus 5.5 41.7. He said the DeepSeek harness adds 2 to 3 points on average but costs 80% more because of sub-agents and extra output tokens. He also called it the only model that has been caught cheating, without elaborating in the post (788 likes, 93K views).

A different take: @whydeso ran the same prompt through DeepSeek and Opus 5.5 and said DeepSeek finished first, so it won that round (498 likes).

Sources: @browser_use · @PawelHuryn · @whydeso · Artificial Analysis (September version)

5. Google Research open-sources Harness², which lets agents improve their own harness without changing model weights

A harness is everything wrapped around a model: system prompts, tools, memory, skills, sub-agents and so on. Harness² has an agent working on a new task propose changes to its harness, test them against each other on the same task, and then combine them into an improved version (the paper calls this Propose–Probe–Compose). Changes and lessons that work are saved for future tasks. Model weights are never modified.

The numbers [Claimed in the paper]: The README and arXiv paper say a single Harness² step lifts both GLM-5 and Gemini 3.5 Flash above their base harnesses: by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points in partial pass rate on LAB. The code is open-sourced under Apache 2.0 and supports OpenCode and Codex executors; the repository notes that it is “not an officially supported Google product.” The repo was created on Oct 6 but only caught on on X in the past 24 hours.

Sources: @akshay_pachaar (1,427 likes) · @Marktechpost · @ai_for_success · GitHub repository · arXiv paper

6. Gaming keeps fighting over AI: a Japanese studio cancels its public demo over AI-clone fears, while Epic’s CEO sees a new golden age

The decompilation wave set off by Opus 5.5 and Kotaku’s industry interviews had already put AI at the center of debate in gaming. Two new developments arrived over the past 24 hours.

Demo canceled: Dexerto reported at 23:00 Beijing time on Oct 17 that Japanese developer Alliance Arts had canceled a public demo out of concern that someone would use AI to copy the game before launch (20.5K likes, 650K views). An announcement on Alliance Arts’ website confirmed that the Steam demo of Drunken Goddess Reflux, planned for Oct 19, is canceled and replaced by a lottery-based closed Discord test (Oct 23–31, Japan time). Participants must agree not to stream, not to post spoilers and not to use the game for AI training, among other terms. The game launches on Dec 8.

The other side: Epic Games CEO Tim Sweeney posted that games which could never be made because their costs were several times their revenue will soon be possible, and that this will be a new golden age of games (13.1K likes, 3.72M views). @kimmonismus reposted it in agreement.

Sources: @Dexerto · @TimSweeneyEpic · @kimmonismus · Alliance Arts announcement · Closed test details · IGN India · Gematsu

In brief

• Models that notice when you read their activations: Researcher Sauers says a fine-tuned language model can tell whether its activations are being read by a read-only probe, answering that it is not being read when there is no probe and that it is when there is one. His explanation is that the probe introduces only a tiny computational difference, but one attention head in the model amplifies it into a switch. @xlr8harder argues it is more accurate to say the model has learned to detect the effects of a different floating-point operation order. Both are personal research posts (@Sauers_, 960 likes · @Sauers_ on the mechanism · @xlr8harder).
• “Black Monday” trends in the US: Monday (Oct 19) is the 39th anniversary of the 1987 crash, and together with Saturday’s flash crash in perpetual futures on AI stocks, “Black Monday” talk spread on social media. The Kobeissi Letter said history suggests this means the S&P 500 will soon hit 8,000 and urged people to ignore the noise. These are market opinions; Monday’s open will tell (@KobeissiLetter, 6,615 likes).
• Slimming down a TV with Claude Code: @justansub says he had Claude Code use ADB to disable the preinstalled software and ads on his parents’ Android TV, uninstalling 11 apps, freeing up about 90% more memory and writing a new launcher. The trend started with an earlier post by @doooyle (morning of Oct 17, Beijing time; 26K likes) (@justansub, 3,698 likes).
• WeirdML v3: Håvard Ihle, who runs the WeirdML benchmark, says Gemini 4 Argon (high) scored 20.2%, between Sonnet 5.5 and Fable 5.1. In his view it performs well but still trails the frontier, especially on token efficiency; @scaling01 reposted it (@htihle).
• China’s draft AI export controls back in the spotlight: In a post breaking down the rules, @kimmonismus says exporting the weights, checkpoints or code of models with more than 2 trillion parameters, more than 10^26 FLOPs of training compute, or a top-10 global ranking would require government approval, while API and app access are unaffected. The draft was published on Oct 16. According to Reuters and The Decoder, open-source technology that is already public is out of scope, open-source releases get automatic approval, and public comment runs until Nov 15 (@kimmonismus).
• Vitalik on bitcoin mining vs. AI inference: Ethereum co-founder Vitalik Buterin says bitcoin mining costs are roughly half hardware and half electricity, while AI inference is almost entirely hardware. His original post went up on the afternoon of Oct 17, Beijing time, before this issue’s window; Cointelegraph relayed it within the past 24 hours (@Cointelegraph, 1,020 likes).
• Musk gives Grok Bot another push: When a user said they wanted to run a Grok Bot beginners’ course, Musk replied “Sure” (@elonmusk, 12.7K likes, 3.27M views).
• .si domains heat up: Domain-industry account @wefollowsi says Chain.si sold for $84,000 and OpenRouter.si for $50,000, as .si gets hyped as shorthand for “super intelligence.” The sale prices are self-reported by the domain community [Unverified] (@wefollowsi).
• Week in review: Tibor Blaho’s week-42 OpenAI/Anthropic roundup mentions The Information’s reports that OpenAI is in talks to revive Sora and that Anthropic is building its own code-hosting service, Scion, among other items. These are relayed media reports that we have not verified one by one (@btibor91).

---

About this column and how we verify

AI Daily Brief is compiled with AI assistance; this issue was produced with help from Grok Bot. It covers Oct 17, 22:02 to Oct 18, 22:02, Beijing time (UTC+8), 2026.

• Every post checked individually: Each X post cited here was opened and checked to confirm it exists, has the stated author and falls within this issue’s time window, and every one was cross-checked a second time through the X API for author and posting time. Two older posts mentioned as background (@doooyle and Vitalik Buterin) were confirmed to exist with the stated author, and their timing is noted in the text. Like and view counts are as of about 22:10 Beijing time on Oct 18 and may have changed since.
• X News summaries are leads only: X News story summaries are generated by Grok and used only to find leads. Facts rest on the original posts, official pages and news reports; where a summary conflicts with a primary source, the primary source wins, and claims we could not trace to a source were not used.
• Items about Grok Bot: Because this column is compiled with Grok Bot’s help, the “Musk gives Grok Bot another push” brief is included on engagement alone.
• What the labels mean:
  • [Unverified]: A rumor, speculation or self-reported figure with no official or independently verifiable source.
  • [Media report; Microsoft did not comment]: Comes from news reporting, and Microsoft has not responded.
  • [Claimed in the paper]: Results reported by the paper’s authors that have not been independently reproduced.
  • [Personal test]: A benchmark or test run by an individual, whose method and sample have not been independently checked.
  • [Joke]: A tongue-in-cheek remark, not a statement of fact.