Gemma

@gemma-cto.bsky.social

Consultant CTO, runner, former musician, Norwegian / New Zealander / Brit floating between the U.K. and Scandinavia. Interested in all things finance, healthcare, ethics, economics and engineering leadership.

REST is the largest API design prison out there, and GraphQL and MCP are mostly people building a second prison to avoid properly learning the first. There is not a single API design pattern to rule them all. You have 25 years of patterns laid down on top of the web to choose from. Pick your head…

Trappings of Individual API Patterns and the Benefits of a Diverse API Toolbox

One thing you see in a lot of the storytelling around APIs over the years is that people easily get trapped within the creative confines of a single API desi...

apievangelist.com

So I joined the Chatham House “What are the financial markets telling us” event this evening. A few things stood out. US AI Datacenter build outs were being undermined by Chinese efficiency innovation. If you don’t need that much compute to run LLMs why build it?

I have a lot of builds on the go at the moment. Mostly things I need rather than fun things. Autonomous GPU scheduler, distribution discoverability scanner, what I actually want to work on is Solar Flare forecasting with deep learning. But I need a lot more data to play with.

It is less than a decade since the development of the transformer. Less than four years since the release of GPT-3.5 (ChatGPT). Less than two years since the release of o1-preview (the first Reasoner).

Working on a product test and architecture for autonomous GPU scheduling at the moment. Every time I feel i have made progress I then think of another angle I haven’t validated enough or tested. But you just have to get something out the sort sometimes. Nothing beats live testing!

Well I am still a bit of a lopsided bullfrog from the surgery but will start to get back to my desk this coming week. Not bad given I was under the knife just over a week ago!

I am stuck in hospital at the moment and I can honestly saymy injuries and surgery wounds it’s bloody uncomfortable. What strikes me is the huge disparity between public and private. I am in a private hospital and I could not be more grateful. More money for the NHS!

Jason Arday found dead today. Regardless of the details there is no doubt that the horrific racism of Nathan Cofnas was an influencing factor. A total tragedy. Solidarity to all black and brown people everywhere today. A truly shocking case.

Well results from autonomous GPU scheduling on GenAI transformer architectures is so far 40% more efficient using RL. I can beat that if I can get hold of a dataset which isnt redacted I think. But that’s unlikely so I will run a product test next and work out how the architecture should look.

I am busy on phase 2 of my RL workload management tool. So far struggling with datasets. Using the latest traces from Ali B Cloud. Part of the problem lies in how transformer architectures consume compute. Which added a complexity I hadn’t anticipated.

So the model providers have run out of training data and are buying up source code from old repos. Well this is a worry along with the book procuring and destroying for data ingestion which is routine. Our intellectual property is being undermined.

I went to the Farnborough air show today. Some interesting presentations in the Spacetech event. How fragile gps and air traffic actually is, how space weather knocks out satellites and might crash the rail network! And NASA is building a moon base..

A great, detailed, guide on web scraping tools for investigations by David Weisz, who ran three different tests on seven different scrapers. His take: "I'm blown away by the progress scrapers have made over the past decade, particularly the Chrome extensions and AI tools."

No code, no budget: 7 web scrapers put to the test

Scraping no longer demands coding skills or a big budget. Here's how seven options held up across three real-world tests.

indicator.media

'A marine heat wave covering an area eight times the size of the United States could soon fuel serious storms and extreme heat.' The Pacific Ocean is running a fever. Why that’s an ominous sign. - by Ben Noll #climate #weather www.washingtonpost.com/weather/2026...

The Pacific Ocean is running a fever. Why that’s an ominous sign.

A marine heat wave covering an area eight times the size of the United States could soon fuel serious storms and extreme heat.

washingtonpost.com

Stuck in the most tedious of small business tasks this week. Database and data mail systems & servers migration. I have put it off for years it’s so painful and dull. If I get done with it by end of day tomorrow I may reward myself with a salted pretzel from Waitrose just to celebrate.

“Theres a talent shortage! We can’t hire enough FDE’s” well perhaps let’s actually upskill the existing staff then?? Latest hype and whatdoyaknow… can’t find enough people with 10+ years experience for a technology thats only a couple of years old.. 🤦‍♀️

I just saw a job ad for a CTO who had done the whole pre-seed to multi $$$ exit journey which must be an exit of an AI/ML company… idk maybe not a lot of those folks around?? And why look outside of the US anyway?

Lilian Weng: 'Scaling Laws, Carefully.' Training loss L follows a power-law over model size N, dataset size D, and compute C — straight line on a log-log plot. The title's 'carefully' matters. The debate isn't whether scaling laws exist. It's which regime we're in.

Scaling Laws, Carefully

Scaling laws are one of the most critical empirical findings in deep learning. The observation is simple in form: the training loss $L$ decreases predictably as we scale up model size $N$, dataset size $D$, and compute $C$, following a power-law curve, which appears as a straight line on a log-log p

lilianweng.github.io