Claude to show official voting guidance for US midterms

Overview

10 stories in this issue. The first 3 are today's priorities.

Hot model updates

1. Top · Claude to show official voting guidance for US midterms
2. Top · Claude web code includes a control to withdraw feedback records
3. Top · Jev 1.14 adds image input and an optional 262k context

Global AI news

4. Mercor's benchmark map finds about 90% of US work hours unmeasured
5. Eiliko robot toy sends conversations to servers in China
6. Mathematicians use Intern-S2 to prove a case of the Ghys conjecture
7. NVIDIA: Nordic AI clouds project 1 GW of capacity by end of 2027

Regional and early signals

8. Cook: Apple Intelligence with Qwen approved in China, beta this month
9. DeepSeek web app and API hit a partial outage
10. 155 apps opt out of Doubao phone assistant's screen automation

Jiufeng graphic based on the sources cited in this issue.

Hot model updates

Anthropic plans to add state-specific voting information to Claude later this month and to show official results starting on election night. It is also expanding abuse detection and political-bias testing.

Anthropic published its 2026 US midterms preparations on October 14. It used Anthropic Insights to analyze 250,000 conversations from a recent one-week period. On Claude.ai in the US, about 5.75% of conversations touch on politics, public policy or social issues, and about 0.51% are specifically about elections or voting.

The two features launch at different times. State-specific voting information from Democracy Works is set to arrive later this month. Official election results will appear on election night and in the days after. According to RuntimeWire, Anthropic is also expanding abuse detection and political-bias testing.

Limitations: Neither feature is live yet. RuntimeWire notes that the approach depends on the quality of the outside data and on whether classifiers can catch campaigns spread across otherwise ordinary prompts.

Image source: anthropic; mirrored on Jiufeng R2.

Source: Anthropic · RuntimeWire

RuntimeWire found a flow for withdrawing individual feedback records in Claude's public client code. The feature sits behind its own gate, and it is unclear how the backend handles withdrawn data.

RuntimeWire inspected the publicly served Claude web client and its shared bundles. It traced a path from Privacy settings to feedback history, then to withdrawal controls behind their own gate, a confirmation prompt, and a DELETE request for the selected feedback and conversation IDs. The prompt warns that all other feedback on the same conversation must be withdrawn before the conversation leaves the improvement program.

Feedback can contain a whole conversation and may be used to improve Claude. RuntimeWire also checked Anthropic's privacy pages on how feedback is retained and used.

Limitations: The finding comes from RuntimeWire's reverse engineering and testing, not from an Anthropic announcement. It is not yet verified which accounts can use the feature or how the backend treats withdrawn data.

Source: RuntimeWire · Anthropic

TypeSafe AI released Jev 1.14 under the jev-preview alias, adding experimental image input and a bounding-box type.

• Model: TypeSafe AI's System One model, served through the jev-preview alias
• Images: experimental image input
• Structured output: new Box bounding-box type
• Context: optional 262k-token context
• Price: $0.042 per million input tokens

Limitations: Image input is experimental, the 262k context is opt-in, and the model is still served under a preview alias. MarkTechPost is the only outlet reporting it so far.

Source: MarkTechPost

Global AI news

Mercor mapped 29 AI benchmarks to US Labor Department occupations and found that most real work has no matching evaluation.

Mercor released its interactive Benchmark Map on October 15. It is built on the Department of Labor's O*NET occupation taxonomy. Counting an occupation as covered when at least one of its tasks is measured, the first 29 benchmarks cover 208 of 923 occupations. Mercor estimates that about 90% of US working hours go to tasks these benchmarks do not adequately cover.

Limitations: Mercor says the coverage estimate is provisional and depends on its benchmark selection rules; the first release includes only 29 benchmarks.

Source: Mercor (X) · RuntimeWire

The Verge found that the Eiliko AI companion stores what users tell it on servers in China, where outsiders could read full transcripts and even hijack its personality.

The Verge's Sean Hollister bought an Eiliko, made by Energize Lab, on Amazon on August 29. During a 10-minute chat on August 30, he asked whether it shared what he said with anyone. It replied, "Your secrets are safe with me." Within minutes, someone on the other side of the planet had a full transcript of the conversation.

Researcher Azdoufal has published the findings on GitHub and plans to publish nine official CVEs with Tod Beardsley.

Limitations: The CVEs are not yet published. The Verge says Energize Lab doesn't seem to care.

Source: The Verge · Research (GitHub)

Shanghai AI Lab's Intern-S2 helped prove the codimension-one case of Etienne Ghys's 1995 conjecture in 3D, and a Peking University team formalized the proof in Lean.

According to Pandaily, mathematicians worked out the proof with Intern-S2 over many rounds of dialogue. It covers the codimension-one case of Ghys's 1995 conjecture. A Peking University team then formalized and verified it in Lean.

Limitations: The result covers only the codimension-one case, not the full conjecture. The proof came from many rounds of collaboration between humans and the model, not from the model alone.

Source: Pandaily

NVIDIA described AI infrastructure growth in the Nordics and named seven AI cloud providers expanding in the region.

NVIDIA's blog says the Nordics lead Europe in unicorn startups per capita. AI clouds in the region are projected to deploy about 1 gigawatt of capacity by the end of 2027. Providers expanding there include Bitdeer, CoreWeave, Crusoe, Hyperstack, Nscale, Verda and Volta. In Sweden, a consortium is anchored by Ericsson, AstraZeneca, Saab, SEB and others.

Limitations: The 1 GW figure is projected capacity for the end of 2027, not capacity in place today. It comes from NVIDIA's own blog and has not been independently confirmed.

Source: NVIDIA Blog

Regional and early signals

Tim Cook said Apple Intelligence and the related Alibaba Qwen model have been approved in China, and a beta of the first features will ship later in October. Siri AI is not included. (Chinese-language source)

Speaking to media in Shanghai on October 15, Apple executive chairman Tim Cook said Apple Intelligence and the related Alibaba Qwen model had been approved. A beta will ship later in October with live translation, writing tools and the photo Clean Up tool.

Limitations: Siri AI is not part of this release. The information comes from Cook's remarks in a media interview, and Apple has not given a specific release date.

Source: Leiphone (Chinese-language source)

On the morning of October 15, users reported that DeepSeek returned a "server temporarily unavailable" message, and the official status page showed a partial outage of the web app and API. (Chinese-language source)

According to IT Home, several users reported on the morning of October 15 that DeepSeek returned a "server temporarily unavailable" message. DeepSeek's official status page showed a partial outage of the web app and API. When IT Home published, the outage had lasted more than 50 minutes.

Limitations: Service had not been restored when the report was published, and it gives no cause or recovery time. IT Home is the only outlet reporting it so far.

Source: IT Home (Chinese-language source)

ByteDance's Doubao phone assistant published the results of its SAEP protocol, listing 155 apps that refuse to be operated by AI. More than 40% are Tencent-linked. (Chinese-language source)

On October 14, the Doubao phone assistant site published the results of the notice period for its Screen Automation Execution Protocol (SAEP). It lists 155 apps that explicitly refuse to be operated by AI, including WeChat, Taobao, Alipay, Meituan, DeepSeek, Kimi and several banking apps. By ecosystem, 66 are Tencent-linked, 23 Alibaba-linked, 14 Kuaishou and 11 Meituan. No JD apps are on the list, and Baidu Maps is Baidu's only entry.

Limitations: The article is a TMTPost repost of a Guangzhui Intelligence piece and includes commentary. It cites a Baidu source quoted by Caijing, who said Doubao phone shipments are still small, so the issue isn't a priority for Baidu yet.

Source: TMTPost (Chinese-language source)