AI Daily Brief · Oct 20, 2026 | DeepSeek open-sources its DeepInfer inference engine; a DeepSeek model beats GPT-6 Sol in a cyberattack test; Altman's three endgames for AI

DeepSeek dominated AI conversation over the past 24 hours. It open-sourced DeepInfer, the inference framework that runs its own service, in a post that drew nearly 20,000 likes. On the same day, evaluation firm Vals AI said a new DeepSeek model had outperformed OpenAI’s restricted-access GPT-6 Sol on a multi-step hacking test, setting off a debate about the risks of open models. At OpenAI, The Information reported that Sam Altman walked staff through three possible endgames for how the industry might control AI development. In infrastructure and markets, Business Insider reported that some of Amazon’s GPU racks were parked in warehouses because data centers lacked power, and AI stocks slipped as Wall Street opened Tuesday. Rumors about Fable 5.5, GPT-6.5 and other new models piled up without official confirmation, while Microsoft, Waymo and UK government advisers each made moves in Europe.

Today’s highlights

• DeepSeek open-sources its DeepInfer inference framework under Apache 2.0, with support for Nvidia SM100 and Huawei Ascend 950; it currently supports only DeepSeek-V4.1-Flash and leaves out full serving components such as the KV cache system.
• Vals AI’s preliminary result: DeepSeek V4.1 Flash 1017 beats GPT-6 Sol on a multi-step offensive cybersecurity task at a far lower total cost; the full evaluation is due in the coming weeks.
• The Information reports that Altman described three options to staff: an IAEA-style international body, a CERN-style “Demis plan,” and one company that can “melt everyone else’s GPUs,” which he called terrible.
• Business Insider says some of Amazon’s GPU racks were staged in intermediary warehouses for lack of power, which Amazon calls best practice; about eight minutes into Tuesday’s session the Nasdaq was down 0.45%, and South Korea’s Elice slumped on its debut.

1. DeepSeek open-sources its in-house inference framework DeepInfer: Nvidia SM100 and Huawei Ascend 950 support, Apache 2.0 license; the announcement drew 19.1K likes, the most for any AI post in the past 24 hours

The release comes at a sensitive moment for DeepSeek. The US Department of Homeland Security (DHS) has just published an order in the Federal Register (document 2026-21447) requiring federal civilian agencies to remove DeepSeek products as well as third-party products and services developed using DeepSeek, with national security systems exempt. The same day, The Information reported that DeepSeek is co-designing AI chips with Huawei [Media report].

What DeepSeek says: [Company claim] At 17:30 Beijing time on Oct 20, @deepseek_ai announced in a four-post thread that DeepInfer is “now fully open source” (the first post drew 19.1K likes and 1.45M views). It described DeepInfer as the in-house framework that serves DeepSeek-V4.1-Flash at scale and said it is sharing the stack to help the community lower compute costs and energy use; inference code for future DeepSeek models and ongoing optimizations will also land there. DeepSeek also open-sourced deepseek-recipe, a production token encoding and decoding component aligned with the processing logic behind its official API. Yicai also reported that inference code for future models will be released through DeepInfer.

What the README says: DeepInfer currently supports only DeepSeek-V4.1-Flash and runs on Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. Features include optimized kernels such as FlashMLA, DeepGEMM, DeepEP and TileKernels; attention data parallelism; MoE expert parallelism with EPLB load balancing; DSpark draft tokens with confidence-scheduled verification; bounded replay for sliding-window attention; asynchronous scheduling; structured outputs; and batch-invariant inference and deterministic sampling (both Nvidia-only). The release covers offline inference only: the KV cache system, prefill/decode disaggregation and load balancing needed for a full online service are not included. It is licensed under Apache 2.0, and the README credits vLLM as the original primary foundation, with inspiration from SGLang and TensorRT-LLM.

No independent numbers yet: Quartz, citing Bloomberg, said no outside party has yet measured DeepInfer’s efficiency gains. Figures circulating online for daily token volume, cache hit rates and time to first token do not appear in the README, and this column does not use them.

Reactions: vLLM thanked DeepSeek and said it will draw on DeepInfer to keep improving DeepSeek V4.1 Flash inference (1,069 likes); @tianyi listed the features and linked the repository (2,023 likes); @kimmonismus wrote that “model weights alone don’t make AI cheap to run. But this helps,” noting that it runs on Huawei chips (600 likes); @victor207755822 posted “Open source vs. closed source. Light vs. darkness.” (1,893 likes, a personal view); Polymarket’s news flash drew 1,135 likes.

Sources: @deepseek_ai 1 · @deepseek_ai 2 · @deepseek_ai 3 (deepseek-recipe) · @tianyi · @vllm_project · @kimmonismus · @victor207755822 · @Polymarket · GitHub: DeepInfer · Yicai · 21st Century Business Herald · Quartz (citing Bloomberg) · Federal Register public inspection 2026-21447

2. Vals AI’s preliminary result: DeepSeek V4.1 Flash 1017 beats GPT-6 Sol on a multi-step offensive cyber task [Third-party claim]; an “abliterated” version stirs a safety debate

At 17:08 Beijing time on Oct 20, evaluation firm Vals AI said it has been building a cybersecurity benchmark based on simulated corporate networks, and that its first finding is that DeepSeek V4.1 Flash 1017 beats GPT-6 Sol on its multi-step offensive cybersecurity task. GPT-6 Sol was tested through OpenAI’s Daybreak trusted access program. DeepSeek used far more tokens but is much cheaper per token, so it reached the same outcome at a far lower total price; Vals plans to publish the full evaluation “in the coming weeks” (856 likes, 89K views). In a follow-up post, Vals said offensive capability previously limited to a closed-source, trusted-access model is now available to anyone via the latest DeepSeek model (63 likes).

One data point: According to AlphaSignal, the test used the Pi agent harness, and each of the two models was run three times on fresh copies of the environment, with results averaged. Scenario details, scoring rules and full logs have not been published, and Vals itself calls the result preliminary.

Reactions: @scaling01 said he wouldn’t really call it a win over Sol, since DeepSeek is several times less efficient, but beating Kimi-K3, GLM-5.3 Flash and Grok-4.7 is already impressive (178 likes). @Hesamation wrote that “the same cyber capabilities that OpenAI kept behind a gate got beaten by a Chinese open model you can run locally” (488 likes).

The abliterated version: [Personal test] @skeptrune said he tried an “abliterated” version of DeepSeek V4.1 Flash 1017, with the weights responsible for refusing illegal requests removed, and felt “sick to my stomach.” His example: it makes it trivial to pull guest data from a motel running an old ASP site. He concluded that GPU compute will become a federally regulated resource, and said he used to think lab safety staff were wasting their time but is now grateful for them (913 likes, 131K views). @RayFernando1337 marveled at how fast it runs on two DGX Sparks and wrote, “This is a national security risk” (1,538 likes). Both are personal views.

Sources: @ValsAI 1 · @ValsAI 2 · @scaling01 · @Hesamation · @skeptrune · @RayFernando1337 · AlphaSignal

3. The Information: at an all-hands, Altman described three endgames for AI, including one in which a single company “melts everyone else’s GPUs” [Media report, not confirmed by the company]

The Information reported on Oct 19 (US time), in a story by Stephanie Palazzolo and Amir Efrati, that at OpenAI’s all-hands last week Sam Altman outlined three industry proposals for pacing AI development. The first is an international body modeled on the International Atomic Energy Agency (IAEA) that would enforce safety standards; he appeared to favor this one. The second, the “Demis plan,” named after Google DeepMind’s Demis Hassabis, is a CERN-style arrangement in which the major labs merge into or work through a single body that supplies models to participating countries. The third, the “Eliezer plan,” named after Eliezer Yudkowsky, has one company achieve recursive self-improvement and use its economic power to “melt everyone else’s GPUs.” According to India Today, Altman said, “Obviously I think that one’s terrible,” and included it only for completeness.

Background: India Today noted that in 2023 Altman, Greg Brockman and Ilya Sutskever wrote in favor of an IAEA-like body that would inspect AI systems, require audits and track compute and energy use, while Hassabis has more recently favored a self-regulatory body along the lines of FINRA, the US financial industry regulator. CellCog said the original story is paywalled and that OpenAI has not commented publicly.

Reactions: @Hesamation summarized the report (701 likes, 96.6K views); @kimmonismus did the same (614 likes); @beffjezos asked, “Why does it have to be a Singleton?” and wrote, “I reject these 3 paths being the only way forward” (225 likes, a personal view).

Sources: @steph_palazzolo · @Hesamation · @kimmonismus · @beffjezos · The Information · India Today · CellCog

4. The AI infrastructure bill: BI says some Amazon GPU racks went to warehouses for lack of power; AI stocks slip at Tuesday’s open as South Korea’s Elice slumps on its debut

The BI report: [Media report] Business Insider reported on Oct 19, citing internal emails, that Amazon set up a dedicated process in August for receiving AI server racks early. One email projected about $5 billion of Nvidia GB300 racks arriving in August, then about $3 billion a month “for the foreseeable future.” People familiar with the matter said some AWS data centers had space but not enough power, so equipment was staged in intermediary warehouses, and at least one supply-chain team said it “didn’t have a clear forecasted path” to installing the racks. An Amazon spokesperson called staging finished hardware near deployment sites until power is ready “best business practice” and disputed that the company lacks deployment forecasts. BI also reported that AWS infrastructure chief Prasad Kalyanaraman told an internal meeting in September that the AI business is “hugely free cash flow negative.” Amazon plans to spend $220 billion this year; AWS revenue grew 37% in the second quarter, its backlog stands at $496 billion, and it still expects demand to exceed capacity in 2027.

The debate: @edzitron wrote that Business Insider confirms “Amazon has warehouses of GPUs that can’t be installed” (4,548 likes, 553K views), and later tied it to Amazon’s chip SPV (70 likes). Reuters reported on Oct 2, citing the Financial Times, that Amazon was in talks to move about $8 billion of Nvidia Grace Blackwell chips into a special-purpose vehicle that would lease them back, with no deal reached. On the other side, @PeterBerezinBCA asked, “If there is a shadow glut of GPUs, why are rental rates on older chips still rising?” (241 likes). @iamgingertrash argued that “the datacenters aren’t really going to get built” and that warehouses of GPUs will likely be offloaded in 2027 and 2028 as defaults loom: “Compute is going to be dirt cheap again” (1,086 likes, a personal view). BI attributes the problem to a lack of power at some sites; it does not say demand is weak.

Britain’s power problem: Reuters reported that Microsoft has pulled out of Nscale’s data center project in Loughton, England, saying, “This decision relates to a single site”; grid officials told Nscale that power would not be available until the 2030s. The site was due to start operating early next year with 23,040 Nvidia GB300 GPUs and 50 megawatts of capacity; the BBC said Nscale still plans to build it. @cfdownes_ contrasted the estimated 10GW of AI data-center capacity the US has added over two years with Britain’s flagship 50MW project losing its customer to grid delays (1,267 likes; the 10GW figure is an estimate he cited).

Wall Street’s open: US markets opened at 21:30 Beijing time. Yahoo Finance’s reading at 9:38 a.m. ET, about eight minutes into the session, had the Dow little changed, the S&P 500 down 0.3% and the Nasdaq down 0.45%; these are not closing figures. The recent rebound in AI stocks stalled, partly because South Korean AI data-center company Elice tumbled on its first day of trading in Seoul: Yahoo put the drop at 32%, while Reuters’ premarket report said nearly 30%, after an IPO that raised 201.1 billion won (about $151 million). Reuters said CoreWeave and Nebius each fell 4.9% before the bell and Nvidia about 1%. SpaceX and Netflix report after Tuesday’s close, Meta and Tesla on Wednesday.

Sources: @edzitron 1 · @edzitron 2 (SPV) · @PeterBerezinBCA · @iamgingertrash · @cfdownes_ · Business Insider · Reuters (Oct 2, Amazon SPV) · Reuters via CNA (Microsoft exits Nscale deal) · BBC · Yahoo Finance live · Reuters via Investing.com (premarket)

5. A day of rumors: “Fable 5.5 this Tuesday” doesn’t happen, a “GPT-6.5 countdown site” looks like a prank, and GPT-6.1 Terra and Kimi K3.1 surface in code [Unverified]

The rumors started the day before: The Information reported on Oct 19 that Anthropic is preparing a new model that is better at inferring what users want, that Google is working on Gemini 4 Carbon, and that OpenAI may release a new Terra model in the coming weeks. None of the three companies has confirmed this.

Fable 5.5: @kimmonismus, quoting @ChrisGPT’s “tomorrow,” wrote: “Looks like today is the day: Fable 5.5 coming this tuesday” (1,135 likes). @signulll listed Anthropic’s lineup “right before ipo” as Fable 5.5, Opus 5.5, Sonnet 5.5 and Haiku 5.5 (4,141 likes). @theo said he works hard to hide when he has early access, but Fable 5.1 got confused after calling Opus 5.5 and “hallucinates that it’s Fable 5.5,” leading people to claim his team leaked a model it doesn’t have (1,299 likes). As of 22:05 Beijing time on Oct 20, neither @AnthropicAI nor @claudeai had released Fable 5.5.

Opus 5.5 nerfed? [One-sided claim] @pcshipp posted, “Hey Claude users, bad news Opus 5.5 just got nerfed” (2,309 likes, 425K views), with no evidence. @pjwerneck replied, “Probably a cache miss for you” (35 likes).

The GPT-6.5 site: @TokenGremlin said the opensi.lol domain hides “6.5” in its source code along with a countdown to 10 a.m. San Francisco time on Nov 1 — exactly day 28 of Tibo’s 28 days (1,673 likes, 189K views). @ChrisGPT found that a Sept 23 archive described the same domain as an “independent OpenAI parody” for sale (200 likes), and @synthwavedd pointed out that the site once carried a Stripe donation form (460 likes). TokenGremlin later concluded that the whole “GPT-6.5” thing “was just a fake leak” (86 likes). OrcaRouter, which checked the page, said the hidden 6.5 and the countdown are real but nothing ties the site to OpenAI, and OpenAI has not responded.

New names in code: @testingcatalog said OpenAI seems to be preparing GPT-6.1 Terra, citing a reference to the model in a Microsoft PR for evaluation configs (283 likes); that shows a team plans to call the model, not that it has launched. @AiBattle_ said Kimi K3.1 appeared in two SGLang commits, one of which references a K3.1 deployment on Deepinfra (88 likes).

Sources: @kimmonismus · @signulll · @theo · @pcshipp · @pjwerneck · @TokenGremlin 1 · @ChrisGPT · @synthwavedd · @TokenGremlin 2 · @testingcatalog · @AiBattle_ · OrcaRouter (opensi.lol check)

6. In Europe: Microsoft rolls out sovereign-cloud features, Waymo heads to Madrid for a 2027 launch, and UK advisers warn against becoming a “high-value AI customer”

Microsoft: [Company claim] Satya Nadella posted about new digital-sovereignty capabilities on Oct 20 (652 likes). According to Microsoft’s blog, the Sovereign Control Panel will enter public preview at Ignite; Data Guardian is coming for the EU and EFTA; external key management with Azure Managed HSM is generally available; Azure Confidential Clean Rooms are in limited preview; confidential VMs on Azure Local are generally available; GitHub Enterprise Local and Foundry Local will become generally available at Ignite; and Bleu serves as the partner cloud in France. The same day in Paris, Nadella highlighted how Saint-Gobain is using Microsoft Discovery to develop construction materials (316 likes).

Waymo: [Company claim] Waymo announced it is coming to Spain, starting with the Madrid region, its first in Spain and second in the European Union, with plans to launch fully autonomous ride-hailing there in 2027 (1,586 likes). Its blog says Jaguar I-PACE vehicles will first drive the roads manually. Waymo says it has completed more than 20 million trips and over 400 million autonomous kilometers, with 95% fewer serious-injury crashes than human drivers.

Britain: Reuters reported that the UK’s Council for Science and Technology (CST) wrote to AI minister Kanishka Narayan on Oct 19 warning that without deliberate, urgent action Britain risks becoming a “high-value AI customer” rather than a country that shapes the technology. It recommended a single government lead and exploring a publicly governed British AI model (shared by Polymarket, 326 likes).

Also: Nvidia said Jensen Huang’s GTC Berlin keynote is tomorrow (@NVIDIAAI, 108 likes).

Sources: @satyanadella 1 · @satyanadella 2 (Saint-Gobain) · @Waymo · @Polymarket · @NVIDIAAI · Microsoft blog · Waymo blog · Reuters via CNA (UK advisers’ warning)

In brief

• Anthropic uses AI to design scorpion-toxin binders: An internal research model running in Claude Science worked for about 64 hours and produced 318,484 designs; two contract labs tested 211 of them, and 17 met the binding bar (7 for Cn2, 4 for AaH II, 4 for AaH I, 2 for LqhIII, none for Ts1), about 8.1% (17 ÷ 211, our calculation). Neutralization has not been tested, Anthropic explicitly says this should not be called an antivenom, and the data are released under CC BY 4.0 (@HuggingPapers, 10 likes · Anthropic research blog · Hugging Face dataset) [Company claim].
• Fields medalist John Pardon on AI: In an interview published Oct 12, the 2026 Fields Medal winner said AI is best at proofreading and finding errors and can make sense of even his roughest proof drafts (“The AI understands. 100%.”); it has helped him prove a few nontrivial lemmas but has not yet produced a major new result (@kimmonismus, 351 likes · Future Class interview).
• SemiAnalysis: Vera Rubin up to 7.3x better perf per dollar than GB300: The claim covers 1-million-token-context agentic inference workloads, compared with GB300 NVL72 (@SemiAnalysis_, 437 likes) [Third-party claim].
• Meta Muse works on Confidential VMs: TestingCatalog found the feature in a hidden setting on Muse web and says it will initially be limited to Meta AI subscribers (@testingcatalog, 101 likes) [Unverified].

---

About this column and how we verify

• AI-assisted: AI Daily Brief is compiled with AI assistance; this edition was generated with the help of Grok Bot and covers 22:05 Beijing time on Oct 19, 2026 to 22:05 Beijing time on Oct 20, 2026.
• Every post checked: Each cited post was read individually to confirm it exists, that the author matches and that it was published within this edition’s time range; all cited posts were also re-read through the X API, with authors and timestamps matching. Like and view counts were read at about 22:14 Beijing time on Oct 20. US stock figures are from about eight minutes into trading, not closing prices.
• X News summaries are leads only: X News summaries are generated by Grok and used only to find leads; facts rest on the original posts, official pages and media reports, and where a summary conflicts with a primary source, the primary source wins. Some stories repackaged reports from early October as new; this edition either dates them explicitly or leaves them out.
• Items involving Grok Bot: This column is produced with Grok Bot’s help; no item in this edition involves Grok Bot.
• What the labels mean:
  • [Media report]: from news reports, not confirmed point by point by the parties through their own channels.
  • [Company claim]: figures or statements published by the company itself, not independently verified.
  • [Third-party claim]: results and judgments published by an evaluation, research or analysis firm, not independently reproduced.
  • [Personal test]: a test run by an individual user, with a small sample and uncontrolled conditions.
  • [Media report, not confirmed by the company]: from news reports; the company involved has not publicly confirmed it.
  • [Unverified]: a rumor or lead not confirmed by an official or reliable primary source.
  • [One-sided claim]: one party’s account, with no evidence offered and no confirmation from the other side.