Mar 5, 2026

Daily Briefing

Talent shock at Qwen, ethics fight, efficiency spikes

Leadership departures at Alibaba’s Qwen team collide with a public ethics clash over a defense contract and a burst of data‑efficiency results. Together they signal a week where capability, governance, and deployment pressures are pulling the ecosystem in different directions. simonwillison.nettechcrunch.comqlabs.sh

Today's Pulse

  • Qwen lead Junyang Lin resigns, followed by other core contributors, prompting an internal all‑hands. simonwillison.net
  • Despite turmoil, Qwen 3.5 drew praise for strong coding performance across sizes. simonwillison.net
  • Anthropic’s Dario Amodei accuses a rival’s defense‑deal messaging of “straight up lies,” while the rival says safeguards exist. techcrunch.com
  • NanoGPT Slowrun reports a jump from 2.4x to 5.5x data efficiency on 100M tokens. qlabs.sh
  • Unsloth ships a practical Qwen3.5 fine‑tuning guide with VRAM savings and broad export options. unsloth.ai
  • BMW brings humanoid robots to its Leipzig plant after a Spartanburg pilot and forms a competence center. press.bmwgroup.com
  • A new suite aims to measure technology’s impact on student learning outcomes. openai.com

What It Means

  • Open‑weight momentum can be fragile when key builders exit, risking slower iteration even if current releases perform well. simonwillison.net
  • Procurement choices will hinge on trust, as defense work reignites scrutiny of provider claims and safeguards. techcrunch.com
  • Efficiency gains show headroom from algorithmic ideas, not just bigger datasets. qlabs.sh
  • Factory pilots point to targeted productivity wins in ergonomically tough tasks rather than full‑line replacement. press.bmwgroup.com

Sector Panels

Tools & Platforms

  • Unsloth’s guide lowers the bar to tune Qwen3.5 from 0.8B to 122B, including vision, with exports to GGUF and vLLM. unsloth.ai
  • Axios details automation that supports local reporters and streamlines newsroom workflows at scale. openai.com

Models & Research

  • Qwen 3.5 earned strong notices for coding, especially given leaner resources than peers. simonwillison.net
  • Slowrun’s 5.5x data‑efficiency bump came via epoch‑shuffle, learned value projections, SwiGLU, and ensembling. qlabs.sh
  • An OpenAI preprint says GPT‑5.2 Pro helped derive and verify nonzero graviton tree amplitudes. openai.com

Infra & Policy

  • Anthropic declined Pentagon terms and criticized a rival’s framing of its deal, which the rival defends as safeguarded. techcrunch.com
  • BMW scales humanoid robotics from a US pilot to Germany and launches a Physical AI competence center. press.bmwgroup.com
  • Roboflow is hiring a Security Engineer to harden infrastructure for computer‑vision deployments. roboflow.com
  • A learning‑outcomes suite targets measurable impact in classrooms over time. openai.com

Deep Dive

Qwen’s leadership turbulence lands amid acclaim for its latest releases. The post recounts Junyang Lin’s resignation, reports of additional departures, and an emergency meeting addressed by Alibaba’s CEO, all against a backdrop of the widely praised 3.5 family. The write‑up highlights strong coding performance across sizes and notes the team’s track record of doing more with fewer resources. The situation remains uncertain, with implications for roadmap pace and continuity. 🔍 simonwillison.net

Why this matters: open‑weight releases have catalyzed tooling, fine‑tuning, and edge deployments around them. Unsloth’s end‑to‑end guide for Qwen3.5 shows the surrounding ecosystem’s readiness to adapt models for specific tasks, from single‑GPU 4‑bit workflows to multi‑GPU and MoE paths, plus exports like GGUF and vLLM. If key contributors leave, future innovation cadence could slow even if today’s artifacts continue to perform well in the field. The contrast between strong outputs and organizational risk is the crux. 🧩 unsloth.aisimonwillison.net

What to watch next: whether leadership transitions stabilize the team and whether additional departures follow. The post underscores both the significance of the resignations and the company’s visible response via the all‑hands. Any shift in priorities could ripple through community adoption, fine‑tuning guidance, and downstream integrations already in motion. For practitioners depending on the 3.5 line, the near‑term focus is likely on preserving current performance while monitoring governance signals. 🧭 simonwillison.net

Introducing ChatGPT for Excel and new financial data integrations (openai.com) OpenAI introduces ChatGPT for Excel and new financial app integrations, powered by GPT-5.4 to accelerate modeling, research, and analysis in regulated environments. openai
Reasoning models struggle to control their chains of thought, and that’s good (openai.com) OpenAI introduces CoT-Control and finds reasoning models struggle to control their chains of thought, reinforcing monitorability as an AI safety safeguard. openai
Introducing GPT-5.4 (openai.com) Introducing GPT-5.4, OpenAI’s most most capable and efficient frontier model for professional work, with state-of-the-art coding, computer use, tool search, and 1M-token context. openai
Nvidia PersonaPlex 7B on Apple Silicon: Full-Duplex Speech-to-Speech in Swift (blog.ivan.digital) Nvidia's PersonaPlex 7B model, designed for Apple Silicon, enables full-duplex speech-to-speech capabilities using Swift. This technology allows for local inference with sub-200ms latency, enhancing t… hn
NanoGPT Slowrun: Language Modeling with Limited Data, Infinite Compute (qlabs.sh) NanoGPT Slowrun is an initiative focused on developing data-efficient learning algorithms for language modeling, emphasizing the disparity between the growth of compute and data. The project aims to a… hn
Anthropic CEO calls OpenAI's messaging around military deal 'straight up lies' (techcrunch.com) Dario Amodei, CEO of Anthropic, criticized OpenAI's messaging regarding its military contract with the Department of Defense, labeling it as "straight up lies." In a memo to staff, Amodei expressed co… hn
Something is afoot in the land of Qwen (simonwillison.net) Recent developments in the Qwen project have raised concerns following the resignation of key team members, including lead researcher Junyang Lin. His departure, attributed to a reorganization within… hn
Qwen3.5 Fine-Tuning Guide – Unsloth Documentation (unsloth.ai) The Qwen3.5 Fine-Tuning Guide provides comprehensive instructions for fine-tuning the Qwen3.5 model family, which includes various sizes from 0.8B to 122B. It supports both text and vision fine-tuning… hn
The five AI value models driving business reinvention (openai.com) Five AI value models show how leaders can sequence AI from workforce fluency to process reinvention and build durable business advantage. openai
Introducing the Adoption news channel (openai.com) Practical insights and frameworks to turn AI progress into business advantage openai
Ensuring AI use in education leads to opportunity (openai.com) OpenAI shares new tools, certifications, and measurement resources to help schools and universities close AI capability gaps and expand opportunity. openai
US tech firms pledge at White House to bear costs of energy for datacenters (theguardian.com) Major US tech companies, including Google, Microsoft, Meta, and Amazon, have committed to covering the costs of new electricity generation for their datacenters during a White House event. This initia… hn
AMD will bring its “Ryzen AI” processors to standard desktop PCs for first time (arstechnica.com) AMD has announced its first Ryzen AI desktop processors, the Ryzen AI 400 series, designed for AM5 desktops. These chips include modern CPU and GPU architectures along with neural processing units (NP… hn
Relicensing with AI-Assisted Rewrite (tuananh.net) Relicensing in open source software presents significant challenges, particularly when it involves contributions from multiple developers. A notable case is the Python character encoding detector, cha… hn
BMW Group to deploy humanoid robots in production in Germany for the first time (press.bmwgroup.com) BMW Group is set to introduce humanoid robots into its production processes in Germany for the first time, specifically at the Leipzig plant. This initiative is part of a broader strategy to integrate… hn
Roboflow (YC S20) Is Hiring a Security Engineer for AI Infra (roboflow.com) Roboflow, a leader in computer vision technology, is seeking a Security Engineer to enhance its AI infrastructure. The company offers a range of products, including tools for model deployment, low-cod… hn