One of my least favourite jobs is answering this RFP question: "How are you using AI in your product?" I understand the intent behind the question. As a buyer of software, you want some confidence that your vendor is up to speed on the latest technology. You want some confidence that what you buy now will improve with the times, or at least won't end up obsolete.

Lucky me. Jury service is high up the UK agenda thanks to I'm just a normal punter who has been on a jury three times, so I definitely don't have the big picture, but I have a view and this is my blog so I'm going to share it 😀 Things I thought were good: On each occasion there was one person on the jury who had some useful real-world experience to bring to the conversation and which affected the decision.

The Economist had an article about how, in the olden days, an employer could infer some useful things about a candidate from how they wrote their cover letter. In the age of LLMs, candidates aren't writing cover letters any more - the machine is doing it (paywalled). The conclusion: Employers, having lost a way of telling strong from weak candidates, have cut all new hires’ pay, and more often end up recruiting worse candidates. 

In Feb 2022 I was experimenting with GPT-3. I was curious at the time about what new avenues this might open up for beyond the fine-tuned BERT and RoBERTa models that I was familiar with for classification tasks. I was curious if GPT-3 represented any deeper knowledge that could help us from a product point of view, for example in finding new suppliers for various goods and services.

I've got a neo4j database which was loaded with data via NeoSemantics. It's got about 15 million nodes and 170 million relationships in - so I'm having to get to grips with how these queries perform at scale. Here's one head-scratching story. First some background. NeoSemantics is a way of loading RDF data into Neo4j. Each node is created with a…

One of my least favourite tech startup jobs is R&D tax credits. I've been doing them for 20-odd years now. From a business perspective it's a chance of free money so it's a no-brainer to try and see what you can get. But as a taxpayer I am horrified at how badly the system works. The taxpayer value for money from this scheme today must be dire.

So I recently had another "interesting" bug in some python code that was down to initializing an empty list in a method signature. It's essentially the same problem I wrote about last time. Here is simplified code for shits and giggles: import random class Foo: def bar(self, arr=[]): for x in range(5): arr.append(random.randrange(1,100)) print(arr) v1 = Foo().bar() # New instance of Foo v2 = Foo().bar() # Another new instance of Foo v3 = Foo().bar([]) # This time explicitly passing in empty array v4 = Foo().bar([101,102]) # This time passing in a new array v5 = Foo().bar() # Basic instance of Foo again…

This is an update to an earlier post on levelling the playing field between investors and startup builders. It assumes a good understanding of current tax policy as it applies to startups. My starting point is that both startup builders and startup investors are important if we want to maximise the UK's chances of building more of an entrepreneurial ecosystem.

As part of my syracuse.1145.am side project, today I learnt that I can put a pointless “where” clause in to convince neo4j to use an index. Compare these two cypher queries: profile MAT…

And ended up wasting a lot of time. I wanted to learn about MCP by setting up an MCP server for The end result is at “Auto” because all it needs is an OpenAPI spec and it generates the …

Here is an interesting study from Benjamin Paaßen doing the rounds on LinkedIn after Gary Marcus reposted it: Large Language models (LLMs) do not simulate human psychology.  My first reaction was "well, obviously". So it was fascinating/scary to see that some people think that LLMs can simulate human psychology, to the point that someone else had to investigate whether it's true or not.