The AI Times
EN
Volume I · Daily editionDaily roundup of AI newsMCPMD
EN
THE AI TIMES
AI in practice

From brief to prototype

GitHub shares the experience of a team that used Copilot to build an interface in half an hour. This is a specific example from a hackathon, rather than a promise of the same speed for every project. There are few fresh official announcements this time, so we also follow up with a selection from the previous edition. We revisit the conditions for faster Sol access, AI glasses helping people with visual impairments, and access checks for company documents. In corporate case studies, we distinguish the vendor’s account from independent verification.

Practical use9 Oct 2026

Copilot helped a team build an interface in half an hour

GitHub shows how people use AI to build prototypes at hackathons. Participant Lance described creating an interface with Copilot in 30 minutes.

Interface creation · minutesGitHub · participant experience
Interface with Copilot
30 min

According to Lance, Copilot also connected the interface to a message verification service. This is one team’s experience, rather than a guaranteed time for every project. Teams at the events also built robots and hardware, alongside applications. GitHub offers students free access to Copilot through the Student Developer Pack. GitHub Copilot Changelog

Subscriptions8 Oct 2026

Faster Sol is reserved for more expensive plans

OpenAI offers Ultrafast mode for GPT-6.1 Sol in Codex and ChatGPT Work. Access is available on the USD 500 Pro plan and eligible Enterprise and Edu plans.

The mode speeds up text generation. The announcement in the source pool does not specify a speed multiplier. Codex changelog

AI in everyday life8 Oct 2026

Meta donates AI glasses to people with disabilities

Meta announced a donation of 1,000 Ray-Ban Meta glasses to organizations in Singapore. They can help people with visual impairments read text and recognize objects.

The announcement includes a USD 30,000 contribution to accessibility education. It does not announce the availability of such support in Czechia. Meta Newsroom

Regulation9 Oct 2026

EU investigates cases of loss of control over AI

The EU scientific panel met to discuss the safety risks of advanced AI. For users, this means oversight of systems whose behavior may have wider consequences.

The panel is investigating recent cases of loss of control and has prepared questions for the makers of the models concerned with the EU AI Office. The panel consists of 60 independent experts who advise European and national authorities. It was due to present safety recommendations to the Commission at the meeting, which Henna Virkkunen was expected to attend. European AI Office

Privacy7 Oct 2026

Enterprise assistant checks permissions at query time too

Amazon Quick and Bedrock knowledge bases check access permissions directly at the source at query time. Employees should therefore receive answers only from documents they are allowed to read.

AWS combines an initial selection of documents with a subsequent check of current permissions. This addresses the risk of the copy of permissions in the database becoming outdated. In the Google Drive example, the system first selects relevant passages and then checks permissions with the original service. A change in group membership on Confluence may not trigger an event that would automatically update the old copy of permissions. AWS ML Blog

AI in practice8 Oct 2026

AI helps build robot simulations from instructions

NVIDIA showcased projects in which AI assembles simulations from everyday instructions. For users, this demonstrates work guided by instructions and checks of results; it is not a finished application for home use.

Developers connect models to Omniverse tools for physics, rendering and virtual sensors. They check and correct results as they go. Using the Astra model, Frank DeLise created a humanoid robot simulator in a warehouse with first-person and third-⁠person views. In the Robo Olympics project, a robot cleared an obstacle in 64 out of 100 simulated attempts. NVIDIA AI Blog

Payments and security8 Oct 2026

AI assistant can pay for services within a fixed limit

Amazon Bedrock AgentCore lets AI assistants pay for individual services. What matters for customers is that infrastructure outside the model itself enforces the spending cap.

AWS describes an Incarna deployment with provider BlockRun in which each model use is priced and paid for separately. The customer owns the payment wallet and grants permission to use it. Incarna uses the USDC digital currency on the Base network, and transactions are publicly verifiable. AWS ML Blog

What experts say

Simon Willison

developer, creator of the LLM and Datasette tools

Quoting Felix Rieseberg

Willison shares Felix Rieseberg’s explanation that the new Cowork moves the working environment to the cloud. Each session has its own isolated environment, and the application mediates access to files on the device. The change is intended to address battery drain and work stopping when the laptop is closed.

simonwillison.net
What experts say

Jack Clark

Anthropic co-founder, author of Import AI

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

Clark analyzes a test of whether AI understands what an image merely implies. The results show a substantial gap compared with humans, but Clark also sees capabilities that some considered beyond standard language models. Common errors include missing a crucial detail or incorrectly identifying an object.

Disclosureco-founded Anthropic and also writes about its competitors

importai.substack.com
AI assistants9 Oct 2026

Postman limits the tools offered to reduce AI errors

Postman described how it is building Agent Mode, an assistant for working with interfaces between programs. For everyday users, the case shows why more options do not necessarily make an assistant more reliable.

In tests, errors in selecting actions increased when the model saw more than roughly 40 tools. Newer models reduced the problem but did not eliminate it. The system selects around 15 tools needed for a specific request from more than 170 tools. Actions that change the application’s state require user consent. AWS ML Blog

Enterprise AI9 Oct 2026

Enterprise AI can refresh documents regularly

AWS summarized September changes to Bedrock, including scheduled knowledge refreshes. This can help employees get answers from more up-to-date company documents.

The managed knowledge base now supports scheduling daily, weekly or monthly refreshes of native data sources. New source connectors include ServiceNow, Confluence Data Center, Salesforce and Zendesk. User-managed configuration is available for SharePoint, OneDrive and Confluence, reducing reliance on administrator accounts. AWS ML Blog

Education7 Oct 2026

Six-week program taught non-programmers to build AI

AWS described a program in which people whose main job was not programming built AI prototypes. For everyday users, the approach of learning through their own specific task with a mentor’s support is of interest.

The program lasted six weeks and took around four hours per week. Results are based on participant self-assessments, rather than an independent test. Teams had three to four members with varying experience and worked on a prototype they had to demonstrate live. AWS ML Blog

Models8 Oct 2026

Mellum2.1 can edit and review code

JetBrains released Mellum2.1 for AI coding assistants. What matters for users is the option to run it on their own device, keeping code and data under their control.

The model explores a project, edits files and checks changes. JetBrains attributes the improvements to further training in real working environments. The model has 12 billion parameters, of which 2.5 billion are activated during use, and an Apache 2.0 license. Packages for Ollama and LM Studio were still in preparation at the time of the announcement. JetBrains AI blog

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) explores an application without source code: native binaries, JavaScript and Electron applications, .NET and websites. It provides evidence and limitations for every conclusion, and 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 an additional 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

Claude Managed Agents

Try dynamic workflows in Claude Managed Agents

Difficulty
Advanced
Time
30 minutes

Based on: platform.claude.com

  1. In the agent definition, configure the multiagent block with the multiagent_20261001 type. Subagents and dynamic workflows are then enabled.
  2. Set a session budget. Each running agent consumes tokens, and the budget also covers workflow runs.
  3. In the system prompt, specify when the agent should start a run: an audit, migration, in-depth research or cross-⁠check.
  4. Submit a task with many parts and follow the run by phase. You can read each agent’s thread.
  5. Give the same task to a single agent and compare the errors found, time and token usage.
Tip

A workflow run has a default lifetime of 24 hours, and a single session can have several open at once. When the budget is reached, the run pauses and resumes once it is increased.

Events in CzechiaNext 14 days · until 25 October

  1. Computing evening: Language models – the foundation of today’s AI

    A lecture in the Computing Evenings series on neural language models behind today’s generative AI.

    LecturePrahaprice on website
  2. Courseonlinefree
  3. ConferencePrahaprice on website
  4. AI HotSpot

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

    ConferencePrahaprice on website
  5. Rakathon 2026

    Help change the future of cancer care with AI!

    Česká asociace AI

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

    Join an open community of people who actually use AI.

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

    MeetingBrnofree

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. 53Claude Fable 5.1
  4. 53GPT-⁠6 Astra
  5. 53Gemini 4 Argon
  6. 52GPT-⁠6.1 Sol
  7. 51Claude Opus 5
  8. 50Claude Fable 5
  9. 48Muse Spark 1.3
  10. 48GPT-⁠6 Sol

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. GPT-⁠6.1 Sol0,72 $
  7. Muse Spark 1.31,60 $
  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)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 12, 2026