#MrDNEJ

Artificial intelligence

AI, from the source

AI is my co-pilot to design, build and ship faster. Here I follow what is released, what it changes for our work, and what is being debated, with a link to every source.

Updated on 28 Oct 2026

Latest releases

  • DeepSeek

    DeepSeek releases V4.1 Flash-1017 as open weights

    The Chinese lab's update improves reasoning, coding and agent tasks while producing shorter, cheaper outputs. The weights are on Hugging Face under the MIT licence, so developers can run the model on their own servers.

    Source
  • Google

    Gemini 4 Argon becomes generally available

    After a first phase reserved for vetted cyber defenders, Google opened its new frontier model Gemini 4 Argon on the Gemini API and Google Cloud. It is not available on the API free tier.

    Source
  • Meta

    Meta releases Muse Wonder, its first frontier model

    Built by Meta Superintelligence Labs, Muse Wonder and Muse Wonder Full focus on agents and on "active memory", the ability to retain and use information provided by the user. Muse Wonder is offered through the Meta Model API at $4 per million input tokens.

    Source
  • Mistral AI

    Mistral unveils Mistral Large 4, a 1-trillion-parameter model

    The French lab trained this multimodal model on its own compute (about 4,000 Nvidia GPUs) as part of Europe's push for sovereign AI. Mistral plans to publish the weights after safety testing.

    Source
  • Anthropic

    Anthropic releases the Claude 5.5 family (Opus, Sonnet, Haiku)

    Claude Opus 5.5 launched on 22 September, followed by Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October. At launch, Opus 5.5 topped the Artificial Analysis and Vals AI overall indices, while Sonnet 5.5 kept the same price as its predecessor.

    Source
  • OpenAI

    GPT-6 Sol and Luna: cheaper models with fewer errors

    OpenAI released Sol, aimed at complex tasks such as coding, and Luna, aimed at high-volume tasks such as summarising documents, at half the API price of the previous generation. Luna is also available to Free and Go users of ChatGPT.

    Source
  • OpenAI

    OpenAI launches GPT-6 Astra, its most capable model

    OpenAI presented Astra as its best model for software engineering and computer use, first available to its Daybreak cybersecurity programme before rolling out to paid ChatGPT plans and the API. Researchers raised concerns that its reasoning technique could make chain-of-thought monitoring harder.

    Source

AI in Africa

  • World Bank Group

    World Bank plans shared AI computing for Africa and Latin America

    The new AI Infrastructure and Readiness Initiative would give governments pay-as-you-go access to shared GPUs instead of each building its own data centre. The World Bank notes that sub-Saharan Africa has 18% of the world's population but only 0.6% of global data-centre capacity.

    Source
  • FuturAfric Intelligence Artificielle

    YIRI: an Ivorian AI platform for local languages

    Launched at Hub Ivoire Tech in Abidjan, YIRI combines a social network and an online marketplace where users can post and sell in their own languages, starting with six Ivorian languages including BaoulΓ© and Dioula. Its founders aim to cover 80% of West Africa within three years.

    Source
  • AssemblΓ©e nationale de CΓ΄te d'Ivoire

    CΓ΄te d'Ivoire's National Assembly trains MPs in artificial intelligence

    Speaker Patrick Achi opened a session on using AI for lawmaking and oversight, insisting that elected officials keep the final decision. Minister Djibril Ouattara presented the 2026-2028 digital roadmap (440 projects, over 2,000 public officials to be trained by end-2027).

    Source

AI in logistics

  • C.H. Robinson

    C.H. Robinson deploys AI agents for ocean, air freight and customs

    Quotes that took one to three days now take two to three minutes, and agents handling US customs documents cut staff time by 23%. Ocean booking agents are rolling out from China to Asia in late 2026 and globally in early 2027.

    Source
  • Flexport

    Flexport lets AI agents book and track freight

    Flexport's MCP server lets assistants such as ChatGPT, Claude and Copilot search rates, book freight at negotiated prices and track shipments in natural language. Its Atlas platform added multimodal routing, AI disruption monitoring and per-container emissions estimates.

    Source
  • Djibouti Ports and Free Zones Authority

    Djibouti unveils AI tools for ports and logistics

    At its first national AI forum, Djibouti showcased DPCS Foresight, which predicts when customers will collect containers to cut re-handling, dwell time and congestion, and an AI camera system that detects road defects along transport corridors.

    Source

The debates

AI and jobs

The case for

Productivity gains can create new jobs, and historically the unemployment effect of a productivity boost fades within about two years; only a minority of firms have fully adopted AI so far.

The concerns

Goldman Sachs estimates that up to 9% of workers in developed economies could be displaced during the transition, with layoffs likely concentrated in the next recession.

Source

Regulation: the EU AI Act timeline

The case for

The Digital Omnibus delayed high-risk rules (to December 2027 for areas like hiring, credit and education) because technical standards and national authorities were not ready, giving companies time to comply properly.

The concerns

Critics see the delays and softer wording, such as on AI literacy, as weakening protections for citizens while AI systems are already being deployed.

Source

Energy and water use of data centres

The case for

Investments such as AWS's $1.5 billion and Microsoft's commitments bring AI infrastructure, cloud capacity and jobs to Africa, which currently has very little computing power.

The concerns

In South Africa, civil rights groups are challenging proposed Cape Town data centres that they say would use over 4.4 billion litres of water a year and compete with residents for scarce electricity.

Source

Copyright and training data

The case for

AI developers argue that training on existing works is fair use and that legal summaries such as headnotes should not be protected by copyright.

The concerns

In September 2026 a US appeals court upheld that ROSS Intelligence infringed Thomson Reuters' Westlaw headnotes when training a competing legal tool, ruling the use was not fair use, though it did not decide the question for generative AI.

Source

Articles

My series L'IA et Nous is in preparation. Meanwhile, my reflections are on the blog.

Read the blog