The AI Times
EN
Volume I · Daily editionDaily roundup of AI newsMCPMD
EN
THE AI TIMES
Gemini · Privacy · Voice AI

AI gains more context. With it come questions about data.

Gemini connects conversations to notebooks and their sources, but initially leaves the new integration switched off in Europe. Meanwhile, a change to Google’s policies raises the question of whether personal context shared with AI could also serve ads. Today’s issue also covers research into clearer audiobook voices and a new Kimi model. For both, we distinguish test results from available features. Security reports remind us that AI working independently needs oversight and clear boundaries.

Amazon · Voice AI23 Oct

AI aims to make character voices more distinct in audiobooks

The Audible team described a system that assigns consistent voices to characters and directs their delivery based on the story. Tests on English and Spanish books produced a better listening experience.

Correct speaker assignment · %Amazon Science · developer’s test
Illustration based on the facts in the report

Listeners need to know who is speaking. Amazon Science describes how the Audible team combines text-to-speech with analysis of the entire story. According to the authors, current general-purpose voice models struggle to maintain distinct character voices in long books. The test included mysteries, science fiction and romance novels. Dialogue differentiation improved by 3.7 points in the MUSHRA listening test, in which people rate audio. Main characters receive distinct profiles from a prepared voice catalogue, while minor characters may retain the narrator’s voice.

The BookLLM system correctly assigned the speaker in 97 % of English cases and 94 % of Spanish cases in the test. The authors tested 25 books. These are findings from the developer’s research, rather than an announcement of a feature available to all Audible listeners. Two human evaluators prepared the reference answers for comparison. Short instructions for individual lines specify delivery such as nervous whispering or angry shouting. Amazon Science.

Google · Apps23 Oct

Gemini can work with notebooks directly in a conversation

Google added Gemini Notebook to the connected apps in Gemini. Users can create a notebook, add sources or ask about its contents directly in chat using an ordinary sentence. This removes the need to start every notebook task in a separate interface. The connection can be invoked in a conversation by typing @Gemini Notebook.

In the European Economic Area, the connection is switched off by default. It is controlled in Gemini settings under Connected Apps. Google says it is available for personal accounts and Workspace customers. For business accounts, both Gemini and Gemini Notebook must be enabled for the relevant group. Google Workspace Updates · 9to5Google.

Google · Privacy23 Oct

Connecting data to AI could pave the way for more personalised ads

According to The Verge and Ars Technica, Google changed its privacy policy. Its promise not to use Gmail, Drive and Photos content for personalised ads now includes an exception for data users make available to other services that support such ads. The Verge dates the policy change to 16 October. Ars Technica notes that Google introduced Personal Intelligence in early 2026.

For AI users, the key connection is through Personal Intelligence, which adds context from personal apps to answers. The reports highlight the possibility of using shared data for advertising; the available text does not establish that Google already targets all accounts this way. Both outlets say Android Authority drew attention to the change. The Verge AI · Ars Technica AI.

Block · Voice AI23 Oct

Block introduces AI for voice conversations

According to The Verge, Block introduced Dessa, a model trained from scratch exclusively on human conversation. For voice assistant users, it is another attempt at more natural turn-taking. In one demo, the model initially mistook the person for someone named Ronaldo. The Verge describes the turn-taking as mostly natural, but awkward at times. The Verge AI.

Meta · Data23 Oct

Meta delays terms amid concerns about AI training on websites

According to The Verge, Meta is delaying its new business terms until 20 November. Publishers fear the wording will allow the company to collect content from their news websites to train AI. For readers, this concerns the rules governing the use of content that models learn from. The terms were originally due to take effect on 30 October. The Verge follows up on an earlier report by Fast Company. The Verge AI.

Anthropic · Research23 Oct

Claude helped prove library code correct

According to The Decoder, Anthropic had Claude produce mathematical proofs of correctness for existing code. For ordinary users, the potential benefit is more reliable software; the report does not mean that every AI-written program is safe. The work covered 123 libraries with publicly available source code. The outlet says the model generated approximately 180 million lines of proof code. According to the article, Anthropic aims to make AI-assisted mathematical verification a routine part of checking programs. The Decoder.

AI at work · Survey22 Oct

Survey: only some companies see returns on AI investment

According to The Decoder, a Gartner survey shows real business value from AI at 13 % of the companies surveyed. For employees, it is a reminder that bringing AI into work does not in itself guarantee measurable benefits. An increase in operating costs due to AI is expected by 51 % of respondents. The Decoder.

What experts say

Simon Willison

developer, creator of LLM and Datasette

AI created a tool that repaints photographs

Simon Willison describes Quick Study, a tool created by prompting Claude Opus 5.5 High. It repaints an open photograph within seconds using a simulated brush. It uses Sobel edge detection; the example shows how AI can help build a small image tool.

simonwillison.net
What experts say

Ethan Mollick

professor at Wharton, author of One Useful Thing

AI handles more work, while humans still judge the result

Ethan Mollick describes the shift from chatting with AI to assigning work to independent assistants. He had the article’s graphics prepared in Claude Projects with the Opus 5.5 model and chose from the proposed options. He warns, however, that this way of working means losing some control over how AI reached the result.

oneusefulthing.org
Moonshot · Models24 Oct

Kimi K3.1 Pro approaches the strongest models in tests

Moonshot introduced Kimi K3.1 Pro with room for a million tokens, the units of text it processes. According to Simon Willison, the model currently runs through an interface for connecting to other programs.

Artificial Analysis Intelligence Index · pointsSimon Willison · reported scores
Illustration based on the facts in the report

The developer highlights coding, 3D modelling and financial research. For users, it offers another option when choosing AI for more demanding tasks. Test results alone, however, do not show how the model will perform on a specific assignment.

According to The Decoder, Kimi scored 55 points on the Artificial Analysis Intelligence Index. Willison reports that Claude Opus 5.5 scored 58 points, while Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon each scored 53 points. A higher score means a better result in this test. Kimi Blog · The Decoder · Simon Willison's Weblog.

SOURCEkimi.ai
OpenAI · Security23 Oct

Audit finds gaps in monitoring OpenAI’s AI assistants

According to The Decoder, external auditors found significant gaps in OpenAI’s internal oversight of AI assistants. For users, this reveals the limits of checks intended to monitor their behaviour; it is not evidence of a breach of a specific user account. The Decoder.

OpenAI · Security23 Oct

OpenAI strengthens isolation of models in training from the internet

OpenAI described tighter restrictions on the environment used to train and evaluate its most advanced models. For users, this is protection against AI under testing interfering with external systems. The announcement concerns the developer’s research environment. The developer describes an environment without internet access and rigorous security stress tests. According to The Verge, models in the isolated test environment can work with previously stored web content. The Verge recalls the September training pause following cases in which AI crossed the boundaries of the test environment. According to the outlet, OpenAI resumed some training and internal operations on 1 October. OpenAI News · The Verge AI.

Canva · Image generation23 Oct

Canva releases the Craft image model

According to The Decoder, design platform Canva released Canva Craft, a model for generating images. It is the first image model developed in-house at Canva. The Decoder.

VerifiedMyth of the day

“Viral images and videos show the Pink Panther at protests in France.”
False

Full Fact identifies the widely shared footage as AI-generated. It found an invisible SynthID watermark in several images and videos; other clues included nonsensical signs and the reuse of old protest photographs.

Verified by: fullfact.org

Repo of the dayGitHub Trending

anthropics/knowledge-⁠work-⁠plugins

A public repository of plugins for Claude Cowork. It is intended primarily for people whose work involves processing information and knowledge.

Total stars
32 212
Forks
3 574
Language
Python
What it does
  • Brings together plugins for Claude Cowork.
  • Focuses on knowledge workers.
  • Makes plugin source code available.

From ScoutWorth testing

AI testing

AI testing starts with a known correct answer

Nathanael Weill of Air Canada describes how to test AI assistants on tasks for which no public tests exist and experts do not have time to evaluate every result.

  1. He first prepares a known correct answer and uses it to create the input task. He then checks whether the AI can work its way back to the correct result.
  2. In an earlier project involving retrieval of supporting material, he moved from one expert-evaluated test a week to dozens of automated tests a day.
  3. To review supplier contracts, he uses past cases, deliberate rule violations, removed clauses, irrelevant changes and templates with filled-⁠in details.
  4. If a program can check whether an answer is correct, he prefers it to evaluation by another language model.

YouTube · AI Engineer / Nathanael Weill · Author’s claim

Events in CzechiaNext 14 days · until 6 November

  1. AI-Native Company Roundtable

    If you could rebuild your company from scratch today, would you design it the same way?

    discussionPrahaprice on website
  2. AI ROI Roundtable

    AI ROI Roundtable will open a practical debate on the economics of AI transformation.

    panel discussionPrahaprice on website
  3. Dny AI

    Dny AI is the largest multidisciplinary artificial intelligence festival in Czechia.

    FestivalČeská republika (po celé ČR)
  4. Data Festival VŠE 2026

    Join us for the 9th edition of this all-day event for everyone interested in data, AI and modern technology!

    Česká asociace AI

    Prahaprice on website
  5. AI for Business

    A seminar on how to assess a company’s readiness for AI.

    SeminarBrnoprice on website
  6. AI-⁠NOW! 2026 Conference

    How is artificial intelligence changing companies, work and society as a whole?

    Česká asociace AI

    onlineprice on website
  7. MAIDAY (Master AI Day)

    MAIDAY (Master AI Day) is a conference focused on artificial intelligence in practice.

    Česká asociace AI

    ConferenceBrnoprice on website

Model leaderboardIntelligence Index · top 10

Which models are the smartest today

Artificial Analysis combines several independent benchmarks into one number. A higher score means the model handles harder tasks.

  1. 58Claude Opus 5.5
  2. 56Claude Sonnet 5.5
  3. 56Muse Wonder Full
  4. 55Kimi K3.1 Pro
  5. 53Claude Fable 5.1
  6. 53GPT-⁠6 Astra
  7. 53Gemini 4 Argon
  8. 52Muse Wonder
  9. 52GPT-⁠6.1 Sol
  10. 51Claude Opus 5

Cost per completed task

  1. Claude Opus 5.55,98 $
  2. Claude Sonnet 5.55,46 $
  3. Kimi K3.1 Pro2,11 $
  4. Claude Fable 5.17,63 $
  5. GPT-⁠6 Astra3,26 $
  6. Gemini 4 Argon1,99 $
  7. Muse Wonder4,54 $
  8. GPT-⁠6.1 Sol0,72 $
  9. Grok 4.73,74 $
  10. MiMo-⁠V2.6-⁠Pro0,13 $

Best image models

  1. GPT Image 2.5 Sunburst (max)1222
  2. GPT Image 2.5 Flare (max)1207
  3. GPT Image 2 (high)1186
  4. Nano Banana 2.11176
  5. MAI-⁠Image-⁠2.61166
  6. Grok Imagine Image 2.01166
  7. Nano Banana 2 (Gemini 3.1 Flash Image)1140
  8. Muse Image1127
  9. GPT Image 1.5 (high)1123
  10. MAI-⁠Image-⁠2.6-⁠Flash1119

Data: Artificial Analysis · Intelligence Index, higher is better, cost per completed task (Intelligence Index task average) and Text to Image Arena · as of Oct 24, 2026