Independent Operator & Newsletter Analysis
Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things (Simonwillison.Net)
Summary: Simon Willison tests Alibaba’s new Qwen 3.8 27B, a 17GB Apache-licensed vision-capable LLM, and finds it remarkably capable—producing excellent bounding boxes, building a custom tool from a single prompt, and driving a coding agent—but crippled by a default ‘xhigh’ reasoning setting that causes spectacular overthinking, like spending 21 minutes and 22,276 reasoning tokens on a pelican SVG. He recommends running it on low or no reasoning for most tasks, and notes that while MTP optimizations can boost speed by ~72%, the model still feels slow at 15-30 tokens/second on consumer hardware. The key takeaway: open-weights models at this size now rival last year’s best proprietary models, but usability depends on taming their defaults.
Why it matters: For independent operators and newsletter analysts, this signals that capable local AI is now a practical tool for content production and tooling—but the default reasoning settings can waste hours of compute, making prompt engineering and configuration choices a real cost factor.
Context: Qwen 3.8 27B follows a rapid cadence of open-weights releases from Alibaba, with the 2.4T-A95B variant arriving just a week earlier. Willison’s pelican benchmark has become a recurring test for local models over the past two years.
"The fact that a 17GB file can do all of this stuff on my home machines is a miracle. Once again, I’m delighted and amazed at how much progress local models have made this year. A year ago this would have been competitive with the best and most expensive of the proprietary models—today it can run on a capable laptop." — SIMONWILLISON.NET
Commentary: Willison’s hands-on testing reveals a practical paradox: the model’s default reasoning setting is a liability, not a feature, and users must actively override it to get usable performance. The 72% speed boost from MTP hints that software optimization, not just hardware, will define the local AI experience. For newsletter operators, the real story is that a $0-cost model can now handle coding, vision, and tool-building tasks that previously required API spend—but only if you’re willing to fiddle with settings and accept slower iteration.
Date: August 16, 2026 06:00 PM ET
URL: https://simonwillison.net/2026/Aug/16/qwen-38-27b/
AI Sentiment Score: Negative (88%)
AI Credibility Score: 10.0/10 — High
Scores and text generated by AI analysis of the source article indicated.Town’s CEO on the self-organizing company (Platformer.News)
Summary: Town, a new AI productivity startup backed by Andreessen Horowitz, is building a self-organizing company by automatically generating personal and team wikis from users’ email, calendar, and documents. CEO Jean-Denis Greze discusses the product’s $49–$59 monthly price, its privacy safeguards, and the imminent launch of a team version that assembles a company knowledge base from individual townies’ data. The interview reveals a deliberate design philosophy that treats AI as a ‘being’ to nurture, while firmly rejecting dark patterns and independent AI economic actors.
Why it matters: For independent operators and newsletter readers, Town represents a concrete shift from manual workspace maintenance to AI-generated organizational infrastructure—potentially eliminating the Notion-style upkeep that many small teams currently endure. The team wiki feature, if it works, could redefine how small companies share knowledge without dedicated admin overhead.
Context: Town emerged from stealth in June 2025 with $55M in funding, and its CEO previously led engineering at Dropbox and served as CTO at Plaid. The product’s ‘townie’ assistant builds a personal dossier within 90 seconds, and the company plans to extend this to a team-level knowledge base that assembles itself from individual users’ data.
"Town’s CEO on the self-organizing company Jean-Denis Greze on AI assistants, building a better corporate workspace, and avoiding "egg on face" This podcast touches on AI. My fiancé works at Anthropic. See." — PLATFORMER.NEWS
Commentary: Greze’s candid admission that the team wiki’s privacy risks could cause ‘egg on face’ is refreshingly honest, but the five-year timeline for trusting LLMs to enforce company privacy policies is optimistic—especially given current hallucination rates. The product’s 15-day data retention and explicit ban on employers reading townie conversations are strong differentiators, but the real test will be whether SMBs accept the trade-off of AI-curated knowledge bases over manual control. For independent operators, the immediate takeaway is that AI-driven organization is moving from novelty to necessity, but the cost and trust barriers remain significant.
Date: August 13, 2026 08:12 PM ET
URL: https://www.platformer.news/town-interview-jean-denis-greze-assistants/
AI Sentiment Score: Negative (54%)
AI Credibility Score: 10.0/10 — High
Scores and text generated by AI analysis of the source article indicated.Post ID: 22d77630