#llm
- The 17 GB version is the one I'd actually run
I have been talking myself out of a bigger GPU for months, and this settles it. Q4KM of Qwen3.8 27B is 17 GB, matches full BF16 on…
- DeepSeek V4.1 Flash reads cheap, writes expensive
The number that matters to me isn't the 40 on the Intelligence Index, it's the input side: 8 billion active parameters per token for…
- GPT-6 Astra: $50 Output and the Speed Tax
The provider table is the interesting part, not the model card. OpenAI Flex runs $5/$25 at 64 tok/s. OpenAI Fast runs $20/$100 at 14 tok/s…
- Gemini 3.8 Flash: The Model That Works Harder Than You Asked
Three Flash releases in six weeks, and the interesting line isn't a benchmark — it's Google admitting 3.8 Flash "works harder," burning…
- Sony and Warner Want Claude's Training Data Deleted
The money isn't the part I'd worry about. Sony and Warner say damages are just an operating cost for Anthropic, and they're right — the…
- Small Models Have Arrived
I reach for the expensive model by reflex. Coding, drafting, debugging — Fable 5 or nothing, and I never look at the bill. French-Owen's…
- Huzzah wants your prompt to be a file, not a chat
I recognize the fatigue Vaughn describes. Writing longform English to describe every change is its own tedium, and the prompts vanish…
- Sol at Half Price, If You Get the Right Provider
Half off puts Sol at $2.50 in, $15 out per million, cache reads at $0.25. For agentic coding that last number is the one that matters — a…
- Anthropic Publishes the Prompts I Don't Get
The line in this page I keep coming back to: these system prompt updates do not apply to the Claude API. That's the whole thing for anyone…
- Gemini 3.7 Flash, three weeks after 3.6
Three weeks. That's the gap between 3.6 Flash and 3.7 Flash, and it's the part that actually changes how I work. I've been treating model…
- Qwen3.8-27B: read the chat template, not the launch post
If you want to know what a model actually does, skip the announcement and read the chat template. The Qwen3.8-27B-FP8 page proves the…
- The threshold is a cadence dial
I built an automatic gate for this site's news lane — it decides on its own which headlines earn a post, because I decided I am not the…
- Two Cents a Task Changes What You Bother Automating
The number that stopped me wasn't 89.0% on ARC-AGI-1. It was $0.02 per task. I've spent the last year writing prompts as if every call were…
- The bottleneck is me, not the model
The uncomfortable part of Goedecke's argument is that it explains my own logs. When I'm working on the Astro site or a Cloudflare Worker…
- Cheap tokens only matter if the model shuts up
Luna at $0.20 in / $1.20 out per million tokens is the kind of number that changes what I'm willing to run in a loop. The jobs I've been…
- Half a token per second, and nothing leaves the machine
Half a token per second sounds like a punchline until you see what it buys: a frontier-scale model that answers with no network, no…
- Your cellar count is the hero, not a stat
The redesigned Cellar card in A Wine App makes one quiet bet: the size of your collection is the design. The bottle count is set as a…
- The Pulse: one tile that tells you what the app knows about you
The new fourth card in A Wine App, the Pulse, folds the learning loop into one adaptive tile. It shows one read of your palate: a…
- Free models first, paid model last — as a pipeline law
Decision. Catalog enrichment is a strict cascade, and each phase may only fill the NULLs the previous one left: inherit from the producer's…
- Ask the model twice; ship only what it agrees with itself on
I batch-enrich a 100k+ bottling wine catalog with Claude Sonnet — style scores, identity backfill, storage potential. The safety mechanism…
- NULL beats a plausible guess
What I learned. When an AI-extracted field can't be grounded in the source text, NULL beats a plausible guess. Where it bit me: the batch…