Anthropic models now available to Microsoft co-pilot users - great! Microsoft shipping data outside its secure cloud to do so, and losing all warranties re residency, copyright protection, etc, into the bargain - not great!

UK emergency messaging test this Sunday 7th at 3pm Anyone who has covert phone may need to take steps to avoid potential discovery, such as those in domestic violence situations educationhub.blog.gov.uk/2025/09/what...

The two sides of living in a seaside town

Agentic browser plugins - very tempting, but also potentially very risky. Especially in healthcare context www.curistica.com/insights/age...

Now is an absolute golden age for Product Engineers.

If your team already has a PM, encourage them to be a vibe coder. If not, hire one who can vibe code. A working prototype is invaluable for gathering user feedback. If you are a PM, learn to vibe code.

Is Chain-of-thought and “reasoning” merely a simulation of reasoning, and therefore constrained by pre-training data? arxiv.org/pdf/2508.01191

Can confirm

I’m about 18 months into learning the alto saxophone, having never played an instrument or studied music before. It’s been a total joy - sharing because if you are wondering if and when you should learn to play something: don’t. Go for it.

A Trust in Cornwall said to their board that cyber security of health systems underwrites #patientsafety. Absolutely and an avoidable death sadly proves that 😥

For all me AI savvy followers, I’d welcome any links or refs to SotA eval approaches for translation using LLMs. It’s an area that seems to be a little stuck, but I could be looking in the wrong places

In September, 2024, physicians working with AI did better at the Healthbench doctor benchmark than either AI or physicians alone. With the release of o3 and GPT-4.1, AI answers are no longer improved on by physicians Error rates appear to be dropping for newer AI models. openai.com/index/health...

I’m new to the MCP game, but running Claude desktop with custom extensions has really opened my eyes to what can be achieved. Connected Notion, Gmail, and WhatsApp in short order. Wow

I’ve built a automated eval platform in 2 hours today. Something that would have been well beyond my capabilities pre-copilot, and even with help would have taken weeks. Incredible

Soggy walk with Maneesh Juneja, still a great way to start Easter Sunday

This is quite an affecting piece of writing, and something that will stay with me for a while. It speaks to that which we individually have control over, which is also nice in current times

It’s an incredible time to be building a startup

"I don't want a Disney vacation of our history! I don't a whitewashed history, I don't want a homogenized history. Tell me the wretched truth about America, because that speaks to our greatness" -- 20 hours into his speech, Cory Booker is spitting absolute 🔥

So it looks like there's a third scaling law: you can make models better by (1) training them with more compute, by (2) having them "think" for longer about an answer, or by (now 3) generating large numbers of answers in parallel & picking good ones Both 2 & 3 seem to have lots of low-hanging fruit

This is a keen observation. Specific prompting techniques have become less valuable as the tech develops, but familiarity with the tools and the ability to clearly describe tasks is taking over

The incredible Minack theatre, about 5 miles from Lands End

Watergate Bay in Cornwall.

I present this as an example of how humans can still misstep when dealing with AI. Particularly when it comes to naming conventions

This is brilliant/spooky: www.threads.net/@evolving.ai...

Have your say on copyright & AI! 📢 The UK government is seeking input on how AI interacts with creative rights. Share your thoughts before the deadline. 🔗 bit.ly/3EWXpnu #AI #Copyright #Consultation

I never thought much about perplexity (I have a hard time spelling their name), but their new "Perplexity Deep Research" is really good and free, within limits. www.perplexity.ai/hub/blog/int...

Another paper showing that humans are poor at discerning between AI and human answers in a clinical setting. Additionally, AI answers perceived more empathetic. What is interesting is that the reason may lie in how the language is used, not necessarily the content

Such a great idea highlighted by @hetanshah.bsky.social - using AI to help people write letters to get what they are entitled to. 1 in 5 people are poor. Dealing with bureaucratic support systems can be huge barrier. #REALChallenge

Polls begin #Realchallenge

Hetan Shah delivering his keynote - “What Ai might mean for our health” 4 themes: Trust, public engagement, market shape, and thinly widely about health #realchallenge @healthfoundation.bsky.social

With all the excitement around #AI, what might it mean for our health? In our fifth #REALChallenge lecture, @hetanshah.bsky.social will warn of the risk of rushing breathlessly to unleash technology. Chaired by Anita Charlesworth. #AIHealthcare https://buff.ly/4j5xhq7

TODAY: Join us for our fifth #REALChallenge lecture. @hetanshah.bsky.social will be exploring what #AI might mean for health, and how to navigate and assess the opportunities and risks. Register now to attend online ⬇️   https://buff.ly/42TRHgu

Hype, Hallucination and Hope - what AI might mean for Health. First time I’ve heard the housekeeping specifically as for #Realchallenge hashtag on Bluesky Heartily support

“…and we go now to the US, where we’re getting images of the FDA website being wiped clean following the President’s executive orders’

What work is underway to upload a mirror?

Apparently, the man who was shot repeatedly with an upholstery gun, is now fully recovered. #dadjokes

Primers • DeepSeek R1 by @amanchadha.bsky.social This primer dissects the innovative architecture behind DeepSeek-R1. aman.ai/primers/ai/d...

Is AI generated content the greenhouse gas equivalent of the possible collapse of the internet as an information resource?

I know that face.....in this week's @bmj.com Tis I!! If any medical people want to join our mini discord here is the link discord.gg/KzSK8SsV

I recommend that people start using DeepSeek Chat with the "DeepThink (R1)" option enabled. It's truly fascinating and interesting to observe this LLM's thought process. It helps that it is free to use. chat.deepseek.com

Fab evening introducing AI to the team at Valentine Voice Care. We have to spend time understanding the basics, as it unlocks rationale use and discourse , and dispels fear and blind trust.

was mostly keeping it together until I read Kyle MacLachlan's tribute to David Lynch

David Lynch has left, and the world is that little bet less weird and wonderful.

I consume a lot of papers about LLMs and their clinical uses, but they often have poor promoting methodology. Here we have an exception - a great paper, interesting findings, and solid approach to model instruction to help support the findings: www.nature.com/articles/s41... #HealthTech #AI #LLM

If the lack of a vitamin is bad, and the right amount of a vitamin is good, always ALWAYS remember that a lot of a vitamin can be bad. This is why doctors will (mostly) be very critical of horse doses of vitamins. Most of the time you pee them out. Sometimes they thrash your nerves.

Current thoughts about AGI: We won’t recognise that it has arrived, or will perceive its first interactions with us as evidence that performance is deteriorating. For many, AGI seems mostly to be defined by user experience, not any objective measure of accuracy or capability.

Oh God, Balatro is on iOS

I’ve stumbled across a game called “Balatro”, widely available on gaming platforms. It’s a cross between solitaire, poker, and Magic the Gathering, and it is brilliant. It does consume your mind like few other games though

In the discussion about AI & AGI, I see next to nothing about Artificial Emotional Intelligence. Is anyone aware of research or work here? I have seen plenty of ability of AI to detect emotional state via proxy measures, and provide emotive/empathic responses, but little on actual emotion reasoning

Clear opportunity to use LLMs to engage trolls right here. If they’re reacting to you, they’re not attacking someone else.