Same run, two answers: a p99 of 10 ms, and a service that stopped responding for half a second. This post builds that fixture in about forty lines of standard-library Python, shows the correction t...
In one paragraph: On 18 September 2026, Vambo AI, a South African startup founded in April 2023, released MORENA, an Apache-2.0 language model trained from scratch for twelve African languages plus...
A score of 0.90 means nine out of ten, but only if the model was calibrated. Most production classifiers are not. This post builds the diagnostic, shows you the two numbers to read before ECE, and ...
The week in AI September 29 to October 5 was the week frontier AI’s finances went public while its rules stayed voluntary. Anthropic’s draft S-1 leaked with a $42bn annual loss inside it, six labs...
A funding round that is really an architecture decision On 15 September 2026, the Egyptian AI company Synapse Analytics closed a US$13 million Series A led by Paris-headquartered Partech, with Alg...
Introduction Prefix caching is the cheapest optimisation in LLM serving and the easiest to lose by accident. Nothing about the model changes; the bytes of your prompt change order, and the input b...
A classroom in northern Kenya, and the number the demo rests on On 21 September 2026, on the Unstoppable Africa mainstage in New York, a Kenyan founder did something more useful than unveil a mode...
Guess ahead, pay once Speculative decoding is the rare inference trick that promises something for nothing: the same output distribution as the model you already run, produced with fewer sequentia...
Introduction The release, in one line On 25 September 2026 the World Health Organization announced that it is expanding its prequalification programme for medical devices to cover computer-aide...
The one-line version Near-duplicate rows double-count your corpus, inflate your evaluation numbers, and make memorised text more likely. MinHash + LSH finds them while comparing a fraction of one p...