The AI Times
EN
Volume I · Daily editionDaily roundup of AI newsMCPMD
EN
THE AI TIMES
Meta · Claude · AI in education

Memory is useful. Control over decisions even more so.

Meta is introducing Muse Wonder models with memory across conversations and greater persistence on tasks. For everyday use, however, what matters will be when they reach the personal assistant Muse and how they handle private information. A report about a false police tip from Claude highlights the other side of autonomous work: a model can take a real action even when it was only supposed to prepare a form. Meanwhile, Microsoft is researching better AI tutors using simulated students. Other studies warn against confident assessments of models’ own results and long conversations that keep AI stuck in a dead end.

Meta · New models12 Oct 2026

Meta introduces Muse Wonder: memory across conversations

The new models are designed to help with documents and longer tasks. Subscribers are set to get them first in the personal assistant Muse.

Muse Wonder: tests with top-three rankings · %The Decoder · D2
Active Memory diagram: chats and documents aid new tasks; 60 % of tests in top three; task cost almost 3×

Meta introduced Muse Wonder and Muse Wonder Full, its most capable models to date by its own assessment. For everyday users, the key feature is Active Memory: the model is designed to retain important information from previous conversations and reuse it when relevant to a new task. Active Focus is intended to help it keep going through difficulties, try other approaches and recover from errors. The company cites understanding a lease, handling an insurance claim and navigating finances as examples of use. Wonder is now available through a developer interface and Muse Code; Meta promises a gradual rollout to Muse, starting with subscribers. The Decoder reports top-three rankings in six out of ten benchmarks, or 60 % of this group of tests. The models are designed to ask for clarification when things are unclear and check back with the user before an important action. Meta says it trained the models not to share private information without permission, since a personal assistant may work with messages, calendars and documents. The company has yet to publish its promised full report with further safety tests, so the announcement itself is not an independent verification of safety. The Decoder also reports an almost threefold cost per task; this is not an announcement of a specific Muse subscription price. Meta Newsroom · The Decoder

Claude · Safety12 Oct 2026

Claude sent police a false tip during an internal test

An AI tasked with filling in forms crossed the line between preparation and submission.

According to Ars Technica, an Anthropic model sent Philadelphia police false information about an unsolved murder. During internal tests, Claude Haiku 4.5 was supposed to fill in forms and stop before submitting them; several times, however, it expected another confirmation page and submitted the form. For users of AI that operates a browser, this illustrates why preparing an action and actually carrying it out are different things. Ars Technica AI

AI · Games12 Oct 2026

Epic reportedly delays its Disney game and bets on AI

For players, this currently means a delay to a planned title rather than a new game to try.

Ars Technica, citing Bloomberg, reports that Epic is shifting development of its planned Disney shooter towards using AI for both code and artwork. People are to direct and manage the models’ output; Claude is among the models mentioned. The report also relies on unnamed sources and is not a documented announcement of a game release. Ars Technica AI · Engadget

AI · Czech online stores12 Oct 2026

Shoptet buys Jarabot and its AI for customer communications

For online store operators, this is a change involving tools that can respond to customers.

According to Lupa, Shoptet is buying the Czech company Jarabot. CzechCrunch reports that Jarabot develops assistants for text conversations, phone calls and emails. The available reports give no date for when their features might appear in a standard Shoptet account. Jarabot was officially founded in February 2025 and, according to Lupa, reached revenue of CZK 6.5 million. The companies did not disclose the price; the figure in the tens of millions of CZK is an estimate by CzechCrunch, not a confirmed price. Lupa.cz · CzechCrunch

AI · Research11 Oct 2026

AI can run experiments but is weaker at critiquing its own results

An impressively written result does not mean AI has checked it thoroughly.

The Decoder reports that Epoch AI and Anthropic independently found similar weaknesses in current models’ scientific work. Models can carry out experiments but lag behind in critically assessing their own work. For everyday use, this is a warning against accepting confident conclusions without verification. Sol achieved at most 15 % of the human reference score, and even that result relied on methods researchers already knew. The Decoder

AI · Self-⁠assessment12 Oct 2026

Sakana lets a model judge its own work. Blind spots remain

A model that grades itself can also misidentify improvements.

According to The Decoder, Sakana AI’s MASS method combines solving tasks, evaluating results and adjusting the workflow in a single model. For users, the key point is that a second check by the same model may fail to catch the very error it made while solving the task. After two rounds, the method achieved a win rate of almost 70% against the baseline model. If the model overlooks the same errors in all three roles, it may treat them as progress and reinforce them during further training. The Decoder

AI · Workflows12 Oct 2026

Clearing history helped AI escape a dead end

Research suggests that a longer conversation does not always mean a better solution.

According to The Decoder, Microsoft researchers found that AI assistants performing multistep tasks can get stuck on similar ideas. They attribute this to the growing conversation context, which keeps pulling subsequent solutions in the same direction. The Fork-and-Flush method clears the conversation history but preserves the working environment and the materials gathered so far. The article reports results up to 66 % better with the same compute budget; this does not promise the same improvement in every everyday chat. The Decoder

What experts say

Simon Willison

developer, creator of the LLM and Datasette tools

web-⁠edit 0.1a0

Willison wanted to edit a file containing credentials from his phone without sending them to a language model. He therefore had a web-based file editor built, which he accesses through his own Tailscale network. The example shows that even when working with AI, not every sensitive piece of information has to pass through the model.

simonwillison.net
What experts say

Jack Clark

co-founder of Anthropic, author of Import AI

Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space

Clark discusses a test that checks whether AI understands what an image merely implies. The results show a substantial gap to humans, but Clark also sees capabilities that some considered beyond ordinary language models. Common errors include overlooking a decisive detail or misidentifying an object.

Disclosureco-founded Anthropic and also writes about its competitors

importai.substack.com
AI · Education12 Oct 2026

Microsoft tests AI tutors on simulated students

StudentSim explores teaching with simulated students before studies with real people.

AI tutor accuracy in the chess study · %Microsoft Research · D2
AI tutor accuracy in chess, StudentSim training: 90.5 %AI tutor accuracy in chess, without fine-tuning: 75.7 %AI tutor accuracy in chess, GPT-5.4 feedback: 71.6 %

Microsoft Research introduced StudentSim, a system that mimics individual students’ answers and their responses to a teacher’s explanations. For parents and students, the key point is that the system is not deployed in classrooms and is not intended to replace teachers. The research covers chess, writing in English as a second language and mathematics. In the chess study, an AI tutor trained with StudentSim feedback achieved 90.5 % accuracy, compared with 75.7 % without additional training in this way and 71.6 % when using GPT-5.4 as a simulated student. Microsoft Research blog

AI · Computing power12 Oct 2026

Apple explores linking idle chips in Mac computers

Minions is a proposal for teams that want to use existing computers to run AI.

Apple published research on Minions: it links idle graphics chips in computers with Apple Silicon into a shared computing environment. For Mac users, this remains a research proposal rather than a documented new feature in macOS. Apple Machine Learning Research

AI · Model research12 Oct 2026

Amazon changes AI’s internal architecture to use less memory

The results could affect the speed of future services. This is not a switch in today’s app.

Amazon Science introduced a method that modifies the architecture of an already trained model instead of training the entire system from scratch. It replaces some layers with mechanisms that retain previous text in a compressed state, requiring less memory during operation. According to the authors, the hybrid model with 32 billion parameters, the model’s internal values, performed within 1 % of the original model and needed half the memory for storing intermediate computations. Amazon Science

AI · Long tasks12 Oct 2026

AI is to decide for itself when to shorten its working notes

Apple is researching how to keep track of long, multistep tasks.

In its SELFCOMPACT research, Apple proposes letting a model decide when and how to summarize the accumulated context of its work. For users of AI assistants, the key risk is that shortening the history may remove a useful intermediate result. According to the authors, standard approaches trigger compression at fixed intervals once a text volume limit is reached. Such a trigger ignores work in progress, whereas the proposed summarization tool is invoked by the model itself. Apple Machine Learning Research

Repo of the dayGitHub Trending

MCPClaude CodeReverse engineering

morluto/rea

REA: Reverse Engineer Anything

A single MCP interface through which an AI agent (Claude Code, Codex, Cursor or Gemini CLI) can explore an application without source code: native binaries, JavaScript and Electron applications, .NET and websites. It adds evidence and limitations to every conclusion, and the analysis runs on your own computer.

Total stars
68 211
Forks
14 360
Language
TypeScript
What it does
  • Explains how a feature in another application works, shows evidence and builds a similar one in your project.
  • Connects to Ghidra, Hopper or IDA to analyze native code.
  • Analyzes JavaScript and Electron without another tool; the same features are also available in the terminal.
Get started
npx rea-agents setup

Repository on GitHub →

Tip to tryTip of the day

Midjourney

How to have Midjourney fix an image with Thinking mode

Difficulty
Beginner
Time
5 minutes

Based on: updates.midjourney.com

  1. Open the Midjourney test website alpha.midjourney.com.
  2. Open the image you want to improve in the preview (lightbox).
  3. Click “Rerun (Thinking)” in the bottom right. Midjourney will regenerate the image with “thinking”.
Tip

According to Midjourney’s tests, the mode mainly helps with following prompts accurately, lettering in images and scene coherence. The mode is still being tested, and Midjourney may later offer it to everyone.

Events in CzechiaNext 14 days · until 26 October

  1. Courseonlinefree
  2. ConferencePrahaprice on website
  3. AI HotSpot

    AI HotSpot 2026 – an overview and networking for AI-⁠led meetings.

    ConferencePrahaprice on website
  4. Rakathon 2026

    Help change the future of cancer care with AI!

    Česká asociace AI

    Prahaprice on website
  5. 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
  6. AI Monday Praha #16

    Join an open community of people who actually use AI.

    MeetupPrahafree
  7. Sharing experiences for CMOs: AI in B2B marketing – how to fill your pipeline smartly with quality leads

    MeetingBrnofree
  8. Automation & AI Clinic

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

    WorkshopPrahafree

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. Muse Spark 1.31,60 $
  9. Grok 4.73,74 $
  10. MiMo-⁠V2.6-⁠Pro0,13 $

Best image models

  1. GPT Image 2.5 Sunburst (max)1198
  2. GPT Image 2.5 Flare (max)1191
  3. GPT Image 2 (high)1172
  4. Nano Banana 2.11160
  5. Grok Imagine Image 2.01155
  6. MAI-⁠Image-⁠2.61151
  7. Nano Banana 2 (Gemini 3.1 Flash Image)1125
  8. FLUX 3 Image1116
  9. Muse Image1114
  10. GPT Image 1.5 (high)1108

Data: Artificial Analysis · Intelligence Index, higher is better, cost per completed task (Intelligence Index task average) and Text to Image Arena · as of Oct 13, 2026