Models & Benchmarksbrief
How to read an AI announcement critically: a worked example
Five checks to run on any model launch post, applied to Anthropic's Claude Haiku 5.5 announcement from October 7.
· 2 min read
03Blog
Writing & Insights
Thoughts on AI, engineering, and building
Practical writing on AI agents, automation, and the engineering behind them, for the people who run the business.
Models & Benchmarksnews-analysis
IBM's Granite 4.2: Apache-licensed reasoning models, and what the vendor-run scores show
IBM released Granite 4.2 in 3B, 8B and 30B sizes under Apache 2.0. What it adds, what IBM's own benchmark table says, and what a buyer should test.
· 3 min read
Models & Benchmarksnews-analysis
Summer 2026 frontier releases: tiers, effort settings and a falling price floor
OpenAI, Anthropic and Moonshot each shipped a major model in July. What matters for buyers is the tiering, the effort controls and the cost per task.
· 4 min read
Models & Benchmarksbrief
Video generation: 30-second clips with native audio change the planning math
ByteDance says Seedance 2.5 makes 30-second clips with joint audio; Google's Veo 3.1 page cites 8-second clips. All claims are vendor-stated.
· 2 min read
Models & Benchmarksbrief
Multimodal AI: reading documents, images and voice
What 'multimodal' means in practice for business documents, scanned pages, images and recorded audio, with the vendor limits to check.
· 2 min read
Models & Benchmarksbrief
Synthetic data for testing AI systems: when it helps and when it misleads
Generated test cases can cover rare scenarios and bootstrap evals, but they can also mirror the generator's blind spots. A short checklist for using them well.
· 2 min read
Models & Benchmarksnews-analysis
Q1 2026 frontier models compared: what the vendor numbers do and do not tell you
Four flagship releases landed within a month. We compare price, features and vendor-reported results, and why the benchmark numbers are not like-for-like.
· 4 min read
Models & Benchmarksbrief
Small language models: what a 9-billion-parameter model can and cannot do
Independent scores for Alibaba's new sub-10B Qwen3.5 models show real gains and a clear weakness: small models still make things up.
· 2 min read
Models & Benchmarksexplainer
Open-weight versus closed AI models: what you actually get, and what to choose
Open-weight models let you run and modify the model yourself; closed models are rented by API. A plain-English comparison and a decision guide for businesses.
· 3 min read
Models & Benchmarksexplainer
Evals explained: how to test an AI system before you trust it
A plain-English guide to AI evals: test sets, graders, repeated trials and why a 75 percent success rate fails more often than it sounds.
· 4 min read
Models & Benchmarksbrief
Benchmark saturation: why leaderboard scores say less than they used to
AI benchmarks are saturating and scores depend on test setup. Three published examples, and how to read a model announcement without being misled.
· 2 min read
Models & Benchmarksbrief
System cards and model cards: what to read before you adopt a model
Vendors publish system cards and model cards with every major model. Here is which sections matter to a business buyer, and a short checklist for reading them.
· 2 min read