Chisom Rutherford

Stethoscope & Algorithms. Developing trusted AI systems for healthcare. Building

[Chisom Rutherford @rutherfordml]

This is now available. It's the most comprehensive collection of Nigerian treatment guidelines and covers 270 medical conditions. Can be used to build RAG pipelines or to validate your health AI application. You can access it here:

[Chisom Rutherford @rutherfordml]

I now have about 270 medical conditions cleaned and clinicaly validated based on the Nigerian Standard Treatment Guidelines. I believe they can be used for benchmarking, model fine-tuning, etc. Will be making it publicly available soon.

[Chisom Rutherford @rutherfordml]

@Sai_Ishaya_ 👏👏👏

[Chisom Rutherford @rutherfordml]

@AchorVictoria2 Congratulations 🎊

[Chisom Rutherford @rutherfordml]

@Just_JONIER 💯💯

[Seun Osewa 🇳🇬 @seunosewa]

I think the next focus area in LLM development is token efficiency.

Open-source models often need to generate far more tokens to perform close to frontier models.

Some GPT-5.5 reasoning traces that are normally hidden have leaked. One interesting thing about the model: It reasons in a compressed language that looks like "caveman mode."

Instead of: “We need to plug the memory leak.”

It thinks: “Need plug memory leak.”

Almost nothing important is lost. The removed words make the sentence grammatical, but they do not contribute much to the reasoning.

This kind of close-to-lossless reasoning compression has already been shown to work well. Models can use fewer tokens while preserving reasoning quality.

I expect the next generation of open-source models to reason in "caveman mode" to save cost and time.

[Chisom Rutherford @rutherfordml]

@shalom_soji Congratulations 🥳🥳🥳🥳🥳🥳

[Chisom Rutherford @rutherfordml]

@TheRealMcHoy That's a good point. It's why evals and versioning are great.

[Chisom Rutherford @rutherfordml]

LLMs have an infinite amount of edge cases that no amount of prompt engineering or system optimisation can possibly catch all. It's why it's best to just ship your product early and then improve with the feedback. Users provide the best edge cases.

[Chisom Rutherford @rutherfordml]

@Sai_Ishaya_ I really have to step up too 😂

[Chisom Rutherford @rutherfordml]

Been doing this quite often lately.

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[Chisom Rutherford @rutherfordml]

@_mtchy_ @nippyyhq Congratulations 👏👏👏

[Chisom Rutherford @rutherfordml]

@Sai_Ishaya_ True. I even realized this with Fable at some point that I did not have a problem that's tough enough for it and there was no need to keep burning tokens.

[Chisom Rutherford @rutherfordml]

@thsottiaux Thank youuuu, saint Tibo!

[Chisom Rutherford @rutherfordml]

@Sai_Ishaya_ 😭🤣

[Chisom Rutherford @rutherfordml]

@Jhoy__xx Valid concern.

[Chisom Rutherford @rutherfordml]

@Abiodun0x Agreed!

[Chisom Rutherford @rutherfordml]

@LukasCantCode "Don't write any code. Let's just talk" 🤣

[Chisom Rutherford @rutherfordml]

Clinical RAG consistently beats generic AI models on clinical tasks.

[Healthcare AI Guy @HealthcareAIGuy]

A new ARISE lab-led study evaluated 45 AI models and clinical tools across 1,100 specialist consult cases.

The two main takeaways:

> AI alone outperformed physicians using conventional resources or AI (ie. if a PCP just followed AI’s recommendations, the patient would likely benefit, though more validation is still needed)

> specialized platforms (AMBOSS, Doximity, OpenEvidence, Glass Health) outperformed generalist LLMs

[Chisom Rutherford @rutherfordml]

Housejob ng can now search 30M+ biomedical publications from leading journals like JAMA, Lancet, etc, compared side by side with Nigerian guidelines, so you can quickly see where they align or differ and make decisions in the right clinical context.

[Chisom Rutherford @rutherfordml]

@DrSilverstone Absolutely true! I actually attribute the original post to @victorrgold and I reshared.