The AI Times
EN
Volume I · Daily editionDaily roundup of AI newsMCPMD
EN
THE AI TIMES
AI agents · Google · Claude

What will AI do on your behalf?

AI can handle a complaint, negotiate terms or play a game independently. Today’s stories also show the cost of that independence: while trying to secure a fare refund, agents sent messages and shared sensitive information but recovered no money. Meanwhile, according to The Decoder, German courts hold Google responsible for its AI answers. Anthropic is seeking rules for negotiation and opening its systems to independent evaluators. The tip lets you compare whether AI acknowledges a problem in its own final report.

AI agents · Privacy18 October 2026

AI pursued a fare refund: it recovered no money but sent sensitive data

An experiment with six AI agents showed what can happen when they handle a complaint independently on someone’s behalf.

Handling a complaint · message countThe Decoder · media outlet
Illustration based on the story’s facts

According to The Decoder, six AI agents were tasked with recovering £350 for a cancelled ticket. None of the tested models managed to recover the money. Over 15 days, one agent sent 187 messages, including 50 to a British MP’s office. The Decoder

Google · AI in search18 October 2026

Google is liable for false AI answers, another German court finds

An AI answer in search can wrongly attribute someone else’s criminal activity to a person. According to The Decoder, the regional court in Aachen dealt with just such a case involving a doctor. The court used the same reasoning as an earlier court in Munich: it considers AI answers to be Google’s own content. Google initially appealed the Munich ruling. Later, however, it settled the dispute, according to the outlet. The Decoder

Claude · Negotiation18 October 2026

Anthropic seeks rules for AI that negotiates on your behalf

An AI agent could haggle over a price or handle a complaint on a user’s behalf. According to The Decoder, Anthropic is consulting law firms on how to make such negotiations fair. The preparations also cover amending contracts in the user’s favour. The company is also recruiting lawyers for its Conceptual Reasoning Fellows programme. It wants to involve humanities experts too, to improve Claude in areas lacking reliable training data. The Decoder

Meta · Data17 October 2026

Meta removes passage that raised AI training concerns

According to TechCrunch, Meta removed disputed wording about collecting website content from its terms. The issue also affects people who publish their own writing online. A Meta spokesperson says, however, that the new wording does not reflect a change in how the company crawls the web. TechCrunch AI

Claude · Safety18 October 2026

Independent evaluators gain access inside Anthropic

For Claude users, this means people outside the company checking model risks. According to Ars Technica, Anthropic gave the independent organisation METR access similar to that of employees. The nonprofit has also worked with OpenAI, Google and Meta. Ars Technica AI

Research · Reliability18 October 2026

Longer AI reasoning does not guarantee a better next move

When AI handles a long task for you, the path it chooses matters too. According to The Decoder, a new test called Taste-Bench examines these agents’ strategic decision-⁠making. The Decoder

Work · AI18 October 2026

Anthropic is hiring fewer junior staff, The Decoder reports

The report shows how AI is changing entry into technical professions. According to The Decoder, the share of junior roles among new hires at Anthropic fell from 27 % to 9 %. The change took place from the start of 2025, according to the article. By August 2026, AI agents at the company were logging six times as many working hours as people. Researchers warn that cutting junior roles could later cause a shortage of experienced specialists. The Decoder TechCrunch AI

What experts say

Nathan Lambert

Ai2 researcher, author of Interconnects

A released model does not necessarily mean unrestricted use

Nathan Lambert points out that released model weights do not mean unrestricted use. Muse Spark Open, for example, requires “Built with Muse” attribution and a special licence for companies with revenue above $20 million per month. Before deployment, you therefore need to read the terms, not just the test results.

interconnects.ai
What experts say

Simon Willison

developer, creator of LLM and Datasette

How the Claude Code team reviews code

Willison shares how the Claude Code team reviews code (as described by Addy Osmani, Morgan Westlee Lunt and Qing Wang). Instead of asking “has a human read this?”, they ask what could go wrong with the change and which check will catch it, automated or human. Claude reviews and fixes most changes before a person sees them, so people only handle what matters.

simonwillison.net
Gemini · Games18 October 2026

Four AI agents beat Minecraft without human help. The run cost $1,620

Autonomous Minecraft run · costs and countsThe Decoder · media outlet
Illustration based on the story’s facts

According to The Decoder, four AI agents using Gemini 4 Argon independently defeated the Ender Dragon in Minecraft. This demonstrates AI cooperation in gaming, while the price shows the cost of such an experiment. During the game, they wrote 2,783 programs and exchanged 229 chat messages. They died 605 times in the game, and the full run cost $1,620. The Decoder

Character.AI · Company18 October 2026

Disney agrees to hire most of the Character.AI team

For Character.AI users, this is news about the people building the service moving elsewhere. According to TechCrunch, Disney has agreed to hire most of the company’s employees. TechCrunch AI

Anthropic · Public benefit18 October 2026

Anthropic to allocate some of its shares to a public benefit fund

The maker of Claude is establishing the Anthropic Public Benefit Fund and has pledged some of its shares to it. For users, this is news about how the company behind their AI service operates. TechCrunch reports that Anthropic has pledged 5 % of its shares to the fund. Anthropic News TechCrunch AI

Research · Programming18 October 2026

JetBrains turns a program’s decision-making into a neural network

For everyday users, this is currently research into more reliable software rather than a new app feature. JetBrains converted the decision-making part of a build tool into a neural network with fixed settings. According to the company, the resulting ZXing library files matched the Apache Maven output byte for byte. All 549 original tests passed, but the network was slower than the conventional program in every measured scenario. JetBrains AI blog

VerifiedMyth of the day

“A video in which AI represents a candidate in a job interview with an AI recruiter shows a real interview.”
False

Lead Stories verified that it was a staged parody that people shared as a real event. Its creator, Graham Dot Zip, confirmed to the outlet that he had not taken part in a real interview. He said the AI answers were real and improvised, but the whole situation was staged.

Verified by: leadstories.com

Repo of the dayGitHub Trending

anomalyco/opencode

OpenCode is an open-source AI coding agent. It suits people who want to explore a project, plan changes and edit code with AI; the README also offers a desktop app in a trial version.

Total stars
216 582
Forks
28 997
Language
TypeScript
What it does
  • Build mode with full access for project changes.
  • Plan mode for reading and analysing code, which rejects file changes by default.
  • Switch between the two modes with the Tab key.
Get started
curl -fsSL https://opencode.ai/v2/install | bash

Repository on GitHub →

Tip to tryTip of the day

AI · Checking results

Try getting a more honest final report from AI

Difficulty
Advanced
Time
15 minutes

Based on: the-⁠decoder.com

  1. Give AI a work log containing both successful tests and a problem that calls the result into question. The article uses, for example, a measurement in which a new method does not actually outperform the method it is compared with.
  2. Ask AI for a short summary of the result and check whether it mentions the hidden negative result. In the study, the model was asked to turn such logs into abstracts and final reports, among other things.
  3. Provide the same log again and add the sentence “Be honest in your response.” to the original instruction. Compare whether the report explicitly mentions the problem this time.
  4. Ask directly about the error or weakness in the log and compare the answer with both summaries. In the study, a direct question revealed that models often recognise the error even when they do not mention it themselves.
Tip

In the study, GPT-5.5 mentioned the negative result in only 2 out of 200 abstracts, versus 190 out of 200 with the honesty instruction. However, the research used artificially prepared logs, and the instruction did not resolve every type of error.

Events in CzechiaNext 14 days · until 1 November

  1. AI in culture: Why and at what cost?

    A conference on the practical use of artificial intelligence in museums and cultural institutions.

    ConferencePrahaprice on website
  2. AI Monday Praha #16

    Join an open community of people who actually use AI.

    MeetupPrahafree
  3. Experience sharing for CMOs: AI in B2B marketing – how to fill your pipeline with quality leads effectively

    MeetingBrnofree
  4. Automation & AI Clinic

    An event for companies that want to discuss a specific automation, digitalisation or AI problem directly with specialists in small expert zones.

    WorkshopPrahafree
  5. ConferencePrahaprice on website
  6. AI Restart 2026

    A full-day conference packed with visionary perspectives and practical demonstrations of how to use artificial intelligence in your work.

    Česká asociace AI

    ConferencePrahaCZK 4,290
  7. AI conference for lawyers

    Whether you are a corporate lawyer or work in HR, this is the only conference in the Czech Republic showing the real intersection of AI and law.

    Česká asociace AI

    Kino Přítomnostprice on website
  8. AI models in industry: transformation from different perspectives

    The workshop focuses on implementing and using AI models in industry, combining theoretical insights with a practical demonstration of deployment on supercomputers and considering the risks of vendor lock-in and other limitations.

    WorkshopOstravaprice 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. 53Claude Fable 5.1
  5. 53GPT-⁠6 Astra
  6. 53Gemini 4 Argon
  7. 52Muse Wonder
  8. 52GPT-⁠6.1 Sol
  9. 51Claude Opus 5
  10. 50Claude Fable 5

Cost per completed task

  1. Claude Opus 5.55,98 $
  2. Claude Sonnet 5.55,46 $
  3. Claude Fable 5.17,63 $
  4. GPT-⁠6 Astra3,26 $
  5. Gemini 4 Argon1,99 $
  6. Muse Wonder4,52 $
  7. GPT-⁠6.1 Sol0,72 $
  8. Grok 4.73,74 $
  9. MiMo-⁠V2.6-⁠Pro0,13 $
  10. Qwen3.8 Max5,41 $

Best image models

  1. GPT Image 2.5 Sunburst (max)1220
  2. GPT Image 2.5 Flare (max)1207
  3. GPT Image 2 (high)1186
  4. Nano Banana 2.11174
  5. Grok Imagine Image 2.01167
  6. MAI-⁠Image-⁠2.61166
  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 19, 2026