Bromide Drift - bromide_drift is the latest unknown image model I have come across on Arena.ai.

It appeared around the same time as Copper Lantern, and I have now managed to get a generation from it using the same British high street prompt I have been trying with these unknown models.

I searched for Bromide Drift afterwards and, as usual, found almost nothing.

I recently wrote about DeepSeek changing the model served behind an existing API model name. It has happened again, not just with a model revision, this time with a minor version 4.1 silently replacing 4.

Another new image model code named Copper Lantern - copper_lantern. I spend quite a bit of time experimenting with models on Arena.ai. I particularly enjoy it when an unknown model appears under a codename, giving you the chance to test it without knowing who made it and try to guess based on the style and codename.

I had some really good looking results from it. I searched for Copper Lantern to see whether anyone knew what it was, but as usual found almost nothing.

Today I came across another image model, this one called Red_fern on Arena.ai where I like to spend time experimenting with new, unknown and existing models. Previously I found one called Cypress, the outputs feel similar in generation style. The code names are both botanical, although the Cypress is a conifer.

Again I searched for Red_fern to see whether anyone knew what it was, but found nothing.

Today I came across an image model called Cypress. I spend quite a bit of time experimenting with models on Arena.ai. I particularly enjoy it when an unknown model appears under a codename, giving you the chance to test it without knowing who made it.

I had an interesting result from it. I searched for Cypress to see whether anyone knew what it was, but found almost nothing.

If you were calling DeepSeek Pro through the API yesterday, the 12th of August, then today the same API call is serving you a different model. The identifier deepseek-v4-pro now resolves to deepseek-v4-pro-0813, where yesterday it did not. You probably did not realise you were on a pre-GA build, because nothing on the pricing page tells you. The only clues sit in blog posts and a footnote deep in the API docs.

Fantastic video by Hannah Fry, my favourite mathematician and broadcaster, who along with software engineer Brendan built an AI agent called Cass, gave it a bank card, and let it loose for two weeks to see what it would do.

I used SolveIt to get me through trying to understand the latest 3blue1brown video Reinventing Entropy | Compression is Intelligence Part 1. This is my "raw" dialog of close reading/watching/listening.

I find SolveIt as valuable for writing and clarifying my thoughts as it is for coding. Just as I find SolveIt makes coding more enjoyable, it does the same for writing. I often learn along the way or make connections between concepts I would not have otherwise.

For me, writing with SolveIt is not about asking an LLM to generate text for you. It is about using the dialog to work together to question your drafts, spot inconsistencies and keep your own voice. This creates a piece of writing I believe is better than you would produce on your own.

Recently, I have been drafting short, dystopian science fiction stories in collaboration with AI, mainly to share as thought experiments rather than being literary pieces. I wanted to explore and highlight potential scenarios in years to come where we face a silent cognitive decline, driven by our over-reliance on AI agents that we no longer fully understand how to build or audit. Also, scenarios such as no novel research, even no novel architectures. As I drafted these stories, I kept noticing a strange recurring pattern. No matter which model I used to write, I kept meeting a character named Elara.

Fantastic short interview by Dr Károly Zsolnai-Fehér with Jeff Dean, who amongst many other things co-founded Google Brain and co-created both TensorFlow and MapReduce!

I have been using SolveIt since late 2024 as part of the first cohort of students, and honestly could not recommend it more. After spending more time with the platform, its growing functionality and tools, certain features stand out to me as genuinely unique in AI tooling.

Running a large language model is not just a compute problem. It is increasingly a memory problem, and the KV cache sits at the centre of it.

I asked an LLM to sum some hours from a text string. It got it wrong. I should not have been surprised.

Are agentic coding tools that good?