AI Highlights

Gemini Canvas turns a paper sketch into a preview app

Key Takeaways
  • •Gemini Canvas turns paper sketches into preview apps
  • •Claude adds personal finance amid accusations
  • •DeepSeek Harness hits 1.54M daily users
  • •OpenAI's data center pledge.
November 6, 2026
30 min read
In this article

Overview

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

Hot Model Updates

  1. Top · Gemini Canvas turns a paper sketch into a preview app
  2. Top · Claude adds personal finance as Silvia's CEO accuses Anthropic of fishing for secrets
  3. Top · DeepSeek Harness reaches 1.54 million daily users, plans a plugin marketplace
  4. Mistral brings Large 4 into Studio with managed agents and workflows

Global AI News

  1. OpenAI pledges to pay power costs at its own data center campuses
  2. Open ASR Leaderboard adds Egyptian and Nigerian English test sets
  3. Amp adds a feedback widget: click a page element, get a code change
  4. MIT Technology Review tries a humanoid cleaning robot that is still remotely piloted

Regional & Early Signals

  1. Doubao Work personal plan drops its 5-hour usage cap
  2. Hunyuan details its internal evaluation: 163 expert blind testers, a major release every 2-3 months
  3. Cainiao opens its global warehouses as a training ground for embodied AI
AI signal map for 2026-11-06

Jiufeng graphic based on the sources cited in this issue.

Hot Model Updates

01/11

Gemini Canvas turns a paper sketch into a preview app

Follow Gemini timeline →

Google's walkthrough: photograph a sketch, hand it to Canvas in Gemini to code, then revise the running preview by chat.

On November 5th Google published a walkthrough with three steps:

  • Step 1: photograph a UI sketch drawn on paper;
  • Step 2: open Canvas in Gemini;
  • Step 3: ask the model to read the drawing and generate code. The result is a running preview app that can be revised through chat.

RuntimeWire says the value for founders is a faster prototype loop: a team gets a clickable preview before spending time on specifications, design files or production engineering.

Limitations: This is a usage walkthrough, not a model release. The available reporting does not name the Gemini version, regions, required plan or success rate. The output is a preview prototype for testing ideas, not a production app.

Transform a quick drawing into a live app with Gemini.

Image source: Google.

Source: Original source Google · RuntimeWire

02/11

Claude adds personal finance as Silvia's CEO accuses Anthropic of fishing for secrets

Follow Claude timeline →

Claude gained an account-connected personal finance workflow, and on the same day Silvia CEO Anthony Pompliano accused Anthropic of courting his engineers for technical details.

Anthropic announced Personal Finance in Claude on November 5th. It is a gradual US beta for Pro, Max and Ultra subscribers, with Plaid-linked finance tools.

That day, Anthony Pompliano, chairman and CEO of Silvia, Inc., posted on X that Anthropic had courted Silvia engineers in April for corporate secrets, technical details and future product plans, and only warned the company days before the launch. He also said Silvia has moved to its own model and hardware at far lower cost.

Limitations: The accusations come only from Pompliano's own X post and are one-sided. The finance feature is a gradual US-only beta.

Source: Original source Anthony Pompliano on X · RuntimeWire

03/11

DeepSeek Harness reaches 1.54 million daily users, plans a plugin marketplace

Follow DeepSeek timeline →

Pandaily reports that DeepSeek Harness processes about 92 trillion tokens a day and has some 35,000 third-party plugins.

Reported figures:

  • Daily users: 1.54 million;
  • Daily tokens: about 92 trillion;
  • Third-party plugins: about 35,000;
  • Next: an official plugin marketplace and a plugin author program.

Limitations: Pandaily is the only report so far; no official DeepSeek announcement was found. The marketplace and author program are plans, not launched products.

Source: Single report Pandaily

04/11

Mistral brings Large 4 into Studio with managed agents and workflows

Mistral is putting Large 4 ("Le Chonk") to work through its enterprise platform Studio, which runs in the cloud, in a VPC or on premises.

Mistral says Large 4 is strong at coding, defensive cybersecurity, and reasoning over complex documents and business tasks. Studio is its enterprise AI platform for building agents, workflows and apps connected to company data and systems.

  • New features: managed agents, workflows, enterprise context;
  • Deployment: cloud, VPC, on-premises;
  • Named customers: ABANCA and BNP Paribas, using it to automate tasks.

Limitations: This is Mistral's own product post, so the customer results are company claims. The post gives no new benchmark scores and no pricing.

Source: Original source Mistral AI

Global AI News

05/11

OpenAI pledges to pay power costs at its own data center campuses

Follow OpenAI timeline →

After voters strongly voiced opposition to data center expansion in this week's midterms, OpenAI published a community compact covering power costs, local input and disclosure.

On November 5th OpenAI published a National Community Compact, which covers its future owned-and-operated campuses, and a Compute Campus Policy Blueprint, which proposes rules for the wider industry.

  • Power: OpenAI pays power costs at campuses it owns and operates;
  • Residents: early input from local residents;
  • Reporting: annual public reports;
  • NDAs: no NDAs with public officials, with only narrow protections for legitimate security information, trade secrets, or competitively sensitive information.

Limitations: RuntimeWire notes that the compact covers only OpenAI's own future sites. Whether other companies face the same terms depends on OpenAI's push for binding industry rules.

Source: Original source OpenAI · National Community Compact · The Verge · RuntimeWire

06/11

Open ASR Leaderboard adds Egyptian and Nigerian English test sets

Hugging Face and Liva AI added two private conversational test sets to the Open ASR Leaderboard.

Earlier, the leaderboard added 11 private splits from Appen and DataOceanAI covering five accents, and then one Indian-accented English split with Voice Arena. The two new sets:

DatasetAccentDuration (h)Style
Liva AI Conversational EGEgyptian1.34Spontaneous conversation
Liva AI Conversational NGNigerian1.55Spontaneous conversation

Both transcriptions are punctuated and cased, and keep disfluencies such as stutters and repetitions.

Limitations: Each set has only 1.3 to 1.6 hours of audio, which is small. Because the sets are private, outsiders cannot reproduce the evaluation themselves.

Source: Original source Hugging Face · GitHub

07/11

Amp adds a feedback widget: click a page element, get a code change

Add one script to a site, and you can annotate elements on the live page and send the instructions to the coding agent Amp.

How it works: enable the feedback widget in the Amp project settings, add one script to the website, then visit the site with ?amp appended. Click any element, write what should change, add more comments if needed, and click Send.

The first time, you sign in with Amp and approve access. Amp then starts a new thread in the project with your comments and makes the change. You ship it from the thread like any other change.

Limitations: The widget is hidden by default and appears only with ?amp or after you have used it once. It requires an Amp account, and changes still have to be shipped from the thread by hand.

Source: Original source Amp

08/11

MIT Technology Review tries a humanoid cleaning robot that is still remotely piloted

Tau Robotics offers humanoid-robot apartment cleaning in San Francisco, but for now humans pilot the robots remotely, and each job also collects training data.

The reporter joined the waitlist through an X post and booked under their real name without saying they were a journalist. Operators steer the robots through cameras in the robots' heads and wrists. Every wiped counter and folded T-shirt becomes training data for more autonomous future robots.

Limitations: Tau's long-term goal is autonomous cleaning, and the report says it is not there yet. Tau charges less because the data is valuable to it, so today's pricing does not show what a commercial service would cost.

Source: Single report MIT Technology Review

Regional & Early Signals

09/11

Doubao Work personal plan drops its 5-hour usage cap

From November 6th, the personal plan of Doubao Work keeps only a weekly quota, with no 5-hour cap. (Chinese-language source)

Quota windowBeforeAfter
5-hour windowCapped; wait for resetRemoved
7-day quotaYesKept, per plan

Before the change, users who hit the 5-hour cap had to wait for it to reset even if weekly quota was left. Users can now concentrate usage within the week. The current plan and remaining quota are shown under subscription and quota settings.

Limitations: The reports give no weekly quota figures for each plan, and the change applies only to the personal plan.

Source: Multiple outlets Leiphone (Chinese-language source) · IT Home (Chinese-language source)

10/11

Hunyuan details its internal evaluation: 163 expert blind testers, a major release every 2-3 months

Tencent says Hunyuan relies on expert blind tests and an internal evaluation system covering 9 business areas, in addition to public benchmarks. (Chinese-language source)

At the "Models and MaaS" session of Tencent's Global Digital Ecosystem Summit on October 30th, Tencent described how Hunyuan is evaluated and developed:

  • Evaluation: besides public benchmarks such as Terminal-Bench and DeepSWE and third-party platforms such as Arena, 163 internal heavy users blind-test the model across four scenario types, including software engineering and game development;
  • Cadence: R&D was reorganized in December 2025 and post-training infrastructure was upgraded in January 2026. Since then, a major version ships every two to three months;
  • Cost: Tencent Cloud MaaS GM Zou Hui said the cost per million tokens has fallen from yuan to jiao or even fen.

Limitations: This comes from a write-up of conference talks. No blind-test scores or comparisons with other models were published.

Source: Single report InfoQ China (Chinese-language source)

11/11

Cainiao opens its global warehouses as a training ground for embodied AI

Cainiao is opening its global warehouse network to AI and robotics companies, with more than 100 test scenarios.

Pandaily reports that Cainiao is letting AI and robotics firms train and test in its real warehouses, with 100+ test scenarios. It also launched a three-day overseas delivery service for tech hardware.

Limitations: So far the only account is Cainiao's own. The participating AI and robotics companies have not been independently disclosed.

Source: Single report Pandaily

This digest is compiled by AI from the sources linked in each item and published after automated fact-checking, without article-by-article human review before publication. If anything differs from the original, the original source prevails. How this digest is made, or report an error: About this digest

Generate one yourself with JIUFENG

Create AI images and videos with sound in your browser: write a prompt, pick a model, download the result. 5 free credits on sign-up, no credit card.

Start creating free