AI — 15 posts
Why Sam Altman Dropped Phones for a 5GW Power Grid
An engineering read of Sam Altman's interview: OpenAI's hardware with Jony Ive, the power wall, and the math limits of recursive self-improvement (RSI).
Why Ollama is slow on Macs and how to calculate bandwidth
Why Ollama runs large language models slowly on Apple Silicon unified memory, what changed with the Metal engine, and how to tune bandwidth in a Modelfile.
When is GPT-6 Astra worth $50 per 1M output, 5x Sol?
GPT-6 Astra reprices the whole request past 272K input tokens, and caching splits 20 calls into $22 versus $6. Modes, migration breaks and monitoring, priced.
Meta's first frontier model Wonder keeps Full off the API
Wonder Full beats three rivals on 7 of 14 model card tests but isn't available through the API, where Wonder costs $4 in and $20 out and leads on 4.
Bonsai 2 squeezes a 54GB 27B model into 6GB for laptops
Bonsai 2 27B shrinks 54GB to 5.95GB. We trace the files, matrices, 128-slot groups, rotation, speed and memory, then 14 benchmarks to see where scores dropped.
Jev only judges at $0.042 per 1M input tokens, output free
Who measured Jev's price, multiples, latency and outside tests, and how. Plus rival models, how confidence is computed, and the weaknesses and security report.
DeepSeek V4.1 Flash grows as its KV cache shrinks 4x
V4.1 Flash's one-quarter, one-eighth and 890 bytes each count a different thing. Four of 40 layers make the 890, and the 511GB of weights sit outside it.
Perplexity keeps files on your Mac, sends only search out
Follows the line between what stays on the Mac and what goes to the cloud: Privacy Gate's four verdicts, PII-Tracer by the numbers, first setup.
Pricey model on easy questions? Cut costs with LLM routing
Sending just 13.4% of questions to the strong model reaches 95% quality, per a paper. We priced a 1M-message chatbot, break-even, and four router placements.
HOTOpenAI posted 722 math manuscripts, how many are verified?
OpenAI's math manuscripts went from 722 to 719, and Lean covers 300 of 719. The 372 bundles, 4,000 problems and 242 guides each count a different unit.