The AI Times
EN
Volume I · Daily editionDaily roundup of AI newsMCPMD
EN
THE AI TIMES
Weekend edition · Highlights from Friday

A quiet weekend: highlights from Friday’s edition

Official sources from AI vendors brought no news worth reporting over the weekend. Today’s edition therefore revisits the strongest stories from Friday, October 10. We open with people who do not code yet built their own AI agents in six weeks: programs that carry out tasks on their own, step by step. OpenAI sped up GPT-6.1 Sol, but only for the most expensive plans. Meta is donating a thousand smart glasses to blind people in Singapore, and EU scientists are examining cases where AI escaped human control. Amazon shows how its workplace assistant checks what you are allowed to see with every query, and Postman explains why an agent with fewer tools makes fewer mistakes. In the Verified section, we examine AI-generated videos of impossible jumps on ramps.

AWS · Education7 October

Non-programmers built their own AI agents in six weeks

Amazon Web Services gave sales and operations staff six weeks to build AI agents: programs that carry out tasks on their own, step by step. For everyday users, the takeaway is that building an AI assistant does not require coding skills; suitable tools, a mentor and room to make mistakes helped more than an engineering background.

Amazon Web Services (AWS), Amazon’s cloud division, described its own program for employees who work with AI daily but do not code: salespeople, consultants, analysts and project managers. Of those surveyed, 90% wanted to try building an AI agent in practice. Yet according to AWS, fewer than a fifth had ever used agent-building tools, and they avoided hackathons because they thought they could not compete with engineers.

The program ran for six weeks, with roughly four hours a week. Teams of three to four chose a real problem from their work, received training and then built a working prototype with a mentor’s help. The mentor helped when they got stuck but did not do the work for them. At the end, teams presented their results to a jury that assessed five areas, including business value and presentation quality.

The winning team of four people without engineering backgrounds built WealthWise, a financial adviser made up of five cooperating AI agents: one analyzes the portfolio, while the others handle risk, financial planning, market developments and investment recommendations. It answers complex queries within five seconds. On the first day, the team agreed on five principles, including “learning matters more than winning” and a short daily meeting.

According to surveys before and after the program, the share of participants rating their understanding of agents as strong or expert rose from 27 to 82%. Readiness to look for AI opportunities at work increased from 41 to 85%. These are participants’ self-assessments, however, rather than independent measurements. According to AWS, participants in the gradual program gained three times more practical skills than people in intensive two-⁠day courses. The program expanded from one team to multiple regions and now runs worldwide. The full description is at aws.amazon.com.

OpenAI · ChatGPT · 8 October

GPT-6.1 Sol gets an Ultrafast mode, only for higher-⁠priced plans

OpenAI added an Ultrafast mode to GPT-6.1 Sol that speeds up response writing. It works in Codex, OpenAI’s coding assistant, and in ChatGPT Work.

It is not available with a standard subscription: access is limited to the Pro plan at 500 dollars and selected Enterprise and Edu plans. Source: developers.openai.com.

Meta · Glasses8 October

Meta to donate a thousand AI glasses to people with disabilities in Singapore

Meta announced in Singapore that it will donate 1,000 Ray-Ban Meta AI glasses to an association for people with visual impairments, a guide dog organization, a disability sports association and Gardens by the Bay. The glasses help blind and partially sighted people read text, identify objects and describe their surroundings.

Gardens by the Bay, which already offers sensory tours for blind people, also plans to use the glasses. More at about.fb.com.

EU · Regulation9 October

EU scientific panel examines cases where AI escaped human control

The European Commission convened a special meeting of its scientific panel on AI. The panel is examining recent cases where people lost control of AI and, together with the EU AI Office, prepared questions for the companies that developed the models involved. For everyday users, this means that a European authority also oversees the most powerful models under the AI Act. The panel consists of 60 independent experts who advise the AI Office and national authorities on matters including risk assessment and market surveillance.

The Commission’s report names neither the companies nor the models. At the meeting, the panel was due to submit recommendations to the Commission on the safety of the most powerful models. Commission Vice-President Henna Virkkunen noted that the EU has the world’s first law addressing systemic AI risks. Source: digital-⁠strategy.ec.europa.eu.

Amazon · Privacy7 October

Amazon’s workplace AI assistant checks what you may see with every query

The workplace assistant Amazon Quick answers questions using documents in SharePoint, Google Drive or Confluence. With every query, Amazon now checks directly with the source in real time whether the person asking is allowed to see the document. For employees, this means AI will not show content they cannot access. Previously, permissions were only copied periodically, so between synchronizations the assistant could show a document a person had already lost access to.

The check has two steps. First, the assistant quickly filters out documents based on stored permissions, then verifies the remaining ones directly with the service, such as Google Drive. According to Amazon, Mondelēz International uses Amazon Quick for more than 35,000 employees. Source: aws.amazon.com.

NVIDIA · Simulation8 October

NVIDIA shows simulations built by an AI agent from everyday instructions

NVIDIA engineers gave an AI agent powered by GPT-6 Astra tasks in ordinary language, and the agent assembled complete simulations from Omniverse libraries: a warehouse with a humanoid robot, self-driving car testing and a model of the International Space Station in a browser. For everyday users, this shows that agents can already build complex applications from descriptions, as well as generate text.

People guided the agents, checked results and managed changes. In the Robo Olympics project, a simulated humanoid cleared one obstacle in 64 out of 100 attempts. When comparing simulated and real cameras and lidars, agents using Astra and Claude Fable 5 worked for about three days. Source: blogs.nvidia.com.

Amazon · Payments8 October

AI agents on Amazon pay for services themselves, with an enforced spending limit

Amazon offers AgentCore payments, a service that lets an AI agent pay for services it needs while working, such as a response from another model. Spending is controlled by a cap set by the customer that the agent cannot change, even if someone manipulates it with a malicious prompt. For everyday users, the key point is that the limit sits outside the AI itself. Using the service, Incarna set up payment for each individual query in three days instead of the planned two to three months.

For now, this is a tool for businesses and developers. During testing, agents made over 1,000 payments, each ranging from 0.001 to 0.05 dollars. Payments use USDC, a stablecoin pegged to the dollar, and every payment can be traced publicly. Source: aws.amazon.com.

What experts say

Jack Clark

Anthropic co-founder, author of Import AI

Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economy

Clark discusses Toby Ord’s argument that swarms of AI agents, multiple programs working on one task at the same time, offer a new way to get more out of AI. Four agents used roughly twice as many tokens, or processed text, to achieve the same result, but by working in parallel they can finish in about half the time. Clark sees this as a new factor in the growth of AI capabilities that will be important to watch.

Disclosureco-founded Anthropic and also writes about its competitors

importai.substack.com
What experts say

Nathan Lambert

Ai2 researcher, author of Interconnects

I expect rapid progress but not towards general superintelligence

Lambert expects AI to significantly speed up engineering work on the models themselves over the next few years, particularly training and operation, while the cost of model intelligence falls almost exponentially. In his view, however, this will not fundamentally change the nature of models, and he does not expect general superintelligence to result. He believes cheaper agents are more likely to drive demand for them.

interconnects.ai
Postman · Agents9 October

Postman finds that an AI agent with fewer tools makes fewer mistakes

Postman, an application for testing and documenting APIs, the interfaces through which programs communicate, built Agent Mode for 40 million developers. The team expected model and prompt quality to be the hardest parts, but integrating the agent into an eleven-year-old product proved more difficult. For everyday users, the lesson also applies to Claude or ChatGPT: the more tools and clutter an agent receives at once, the more easily it makes mistakes.

Postman therefore shows the agent only tools relevant to the task: it selects about 15 from more than 170. Missing context caused even more errors than missing tools: the agent needed to know where the user was in the application and what they had open. A detailed analysis is at aws.amazon.com.

JetBrains · Models8 October

JetBrains releases Mellum2.1, a small open model for coding

JetBrains, a maker of developer tools, released Mellum2.1, a small open model that anyone can download and run on their own computer. It is designed for coding agents: it explores code, edits files and checks its own changes. For everyday users, it shows that even a small model on local hardware can do useful work without code or data leaving the computer. The model has 12 billion parameters, the adjustable numbers of which 2.5 billion are active at each step, and is released under the permissive Apache 2.0 license.

Compared with the previous version, its biggest improvement is in autonomous coding. By predicting multiple words at once, it responds to a request about 1.6 times faster and handles almost twice as much text as Qwen3.5-9B under heavy load. It is available to download from Hugging Face, with versions for Ollama and LM Studio expected soon. Source: blog.jetbrains.com.

AWS · Roundup9 October

AWS rounds up September updates: more efficient agents and fresher company documents

Amazon Web Services rounded up September’s updates to Bedrock, its AI platform for businesses. For everyday users, the main point is that workplace AI assistants built on Bedrock can work with fresher documents: knowledge bases can now refresh themselves daily, weekly or monthly.

Businesses also have a wider choice of models. Source: aws.amazon.com.

Apple · Research8 October

Apple proposes a way to reduce generation to a few steps

Apple researchers introduced a method called Normalizing Trajectory Models (NTM). Diffusion models gradually create an output from noise in many small steps; NTM aims to do the same in a few larger steps. For everyday users, this remains research rather than an application feature. According to the authors, existing methods for generation in a few steps give up exact probability calculations, which NTM retains.

The paper was published on Apple’s machine learning research website. Source: machinelearning.apple.com.

VerifiedMyth of the day

“An athlete runs to the top of a giant turquoise ramp and performs a flip. Such ‘parkour’ videos at famous landmarks are circulating on X, TikTok, YouTube and Facebook as real footage.”
False

Fact-checking newsroom Lead Stories found that many of these videos were generated by AI, most likely for clicks. Detectors Sightengine and FauxLens gave one of them a 92 and 93% probability of being AI-generated, and in the footage the athlete’s arm disappears and her bare foot turns into a shoe. Other videos bear the Dola AI watermark or show spectators with their backs to the athlete they are supposedly filming.

Verified by: leadstories.com

Events in CzechiaNext 14 days · until 25 October

  1. Computer Science Evening: Language models – the foundation of today’s AI

    A lecture in the Computer Science Evenings series on neural language models underpinning 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-⁠led 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 practical uses of artificial intelligence in museums and cultural institutions.

    ConferencePrahaprice on website
  7. AI Monday Praha #16

    Join an open community of people who use AI in practice.

    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