rrsp

This is why it’s luck.

Putting more out there will increase the probability of a reward, but it doesn’t guarantee it.

I built a universal live speech translating app.

I’ve been playing around with the Whisper models for a few years now. Last year I had an idea about how to run Whisper Large v3 in real time. That idea became ScribeAI.

Because the quality of transcripts was so high, much higher than I could get with Parakeet, I started to think about how it would serve as a good input for live translation. I played around with this and was surprised by how good the results is, I’ve used it to follow along political speech’s from foreign leaders and other content I’d have just never been able to consume before. You can translate by bringing your own LLM service API key or using the inbuilt Apple Translate models (for a completely offline experience).

https://apps.apple.com/gb/app/scribeai-transcribe-speech/id6...

98% = 2 units of UV reaching the skin

99% = 1 unit of UV reaching the skin

Thus 98% filtering lets in 2x as much as 99% filtering

Why does psychiatry need to have an ‘equivalent’ of a sprained ankle?

Most people recognise a sprained ankle, at least mild ones, as a self limiting illness. An issue with psychiatric diagnoses is that they are often not taken to be self limiting and often become a large part of a patients self image. While sometimes this can be helpful and help inform treatment it can also be harmful and I have seen this harm first hand in patients I see.

Could you recommend any good resources for learning more about microeconomics?

Something being ‘instinctually normal’ does not make it inherently the safest option

I think another explanation is that the Sahm rule came into an effect last week signalling a possible upcoming US recession.

For context Rheumatoid Factor is present in 4% of the healthy population, an even upto 30% in certain populations like native Americans [1]

ANA is positive in 15% of the population [2]

The idea that the tests rule out serious causes is not really correct, people can have seronegative inflammatory arthritis in which case these tests are negative. Peace of mind is dubious as well given that even with a positive test you are still more likely to not have an inflammatory arthritis.

It’s a common misconception that blood tests are binary and provide concrete answers. Sometimes they do. But most blood tests, like many measurements, are far from binary and have a distribution across the normal population, once it passes an arbitrary threshold it does not necessarily mean you have disease X, context (clinical history and examination) are often far more important in making a diagnosis.

Unfortunately some of the population tend to overweight blood tests vs a physicians assessment, I guess they see the former as an objective measure and the latter as subjective. Especially because if you shop around enough clinicians you’ll eventually find one who will say what you want to hear and that one will inevitably be ‘right’.

1. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3845430/ 2. https://rheumatology.org/patients/antinuclear-antibodies-ana...

How do you fix a fissure in the anus?

A: Rectal Glyceryl trinitrate

How do you fix GERD?

A: For the majority of people weight loss and dietary changes will do the job

How do you fix Crohns?

A: You don’t ’fix’ Crohn’s per se as there is no cure. But most people with Crohn’s can be have their disease managed well with -mab medications

In all those conditions you named surgical management is very much at the end of the list and for a minority of patients who do not respond to the above. The first line management in all of these conditions very much does look at ‘fixing the problem’ contrary to what your espousing.

Both the ‘gold’ answer and the model reference a PA and AP view respectively as well as a lateral chest radiograph. The picture only contains a lateral radiograph though.

“- There's a paper from ~6 months ago by some researchers who claimed to get slightly better performance than Whisper vanilla, but their code is not public, and I can't find any blog posts or articles talking about using their work.”

Have you got a link to this paper?

https://docs.mistral.ai/platform/pricing

Pricing has been released too.

Per 1 million output tokens:

Mistral-medium $8

Mistral-small $1.94

gpt-3.5-turbo-1106 $2

gpt-4-1106-preview $30

gpt-4 $60

gpt-4-32k $120

This suggests that they’re reasonably confident that the mistral-medium model is substantially better than gpt3-5

No I missed it, was looking in the methods section as opposed to the results, thank you!

I wonder if they made any effort to check whether the NEJM case studies that this whole study is based are in the PALM-2 training dataset.

Has the same lack of capitalisation as Sam Altman's message, wonder why

What I’m saying is if it is possible to train transformers to achieve AGI, then why hasn’t it happened yet? What’s the limitation that will be overcome in the next 5 years?

I don’t think it’s so clear. The transformer has been available for 6 years, if it were possible to train one to achieve AGI then what’s stopped anyone from doing this that won’t still be the case in 5 years time, given than there’s potentially ?trillions on the table for anyone that does.

We’re 5+ years from the transformer and we’re still using the transformer for the most cutting edge llms. I don’t see what difference another 5 year is going to make unless someone invents something new that can surpass the transformer, and given the amount of money and resources that has been put into AI since 2017 and the lack of innovation since (in terms of fundamental architecture, not things like Lora and Rope) then I’d say the chances are way way lower than 50%.

This field is littered with studies like this, measuring hundreds of different outcomes and then presenting the statistical noise as positive findings.

Key bit: “Three weeks after treatment began, all groups reported similar symptom scores.“

All groups being people who took ensitrelvir 250mg / ensitrelvir 125mg / placebo

Does anyone know if there’s a reason why we’ve not agreed on a standard way of analysing data for any given type of study or type of outcome?

Maybe it’s naive but my intuition would be that if someone were conducting a randomised control trial with a hypothesis that A > B then there should be a known best practice for analysing that data to check the hypothesis. A way that is reproducible, otherwise surely it’ll become a shitfest of people trying multiple methods of analysis and then publishing whichever produced the ‘best’ result.

The context is that guns are much harder to come by in the UK and as such the armed police in the UK are much less likely to be met by an armed individual. So you would expect there to be far fewer times where they were actually required to open fire.

'Most of the layers within each decoder block have names like gpt2_transformer_layer_3d'

As a doctor organising a scan is the path of least resistance, patients almost universally think that a scan = good care, and why would they not? More data is better surely as another commentator has noted.

So we do the scan and find an ovarian cyst, not to be unexpected, the prevalence of an ovarian cyst is widely quoted as anywhere between 8-15% [1]. You tell the patient that you found an ovarian cyst. Naturally she asks if it’s concerning. It’s a simple cyst, so if we use [1] to inform our figures we can tell her that in 1 years time there’s a 50% chance that the cyst will be gone, a 34% chance it will still be there, a 7.5% chance that there will be more than one cyst and a 5.5% chance that there will be a complex cyst. Simple cysts are not thought to be linked with an increased risk of ovarian cancer, but complex ones are.

Now on hearing that there’s a 5.5% chance of finding a complex cyst next year the patient opts for follow up scanning. They of course Google symptoms of ovarian cancer and see that bloating is a symptom. The patient worries, she has very bothersome bloating, she reads about doctors missing ovarian cancer and worries if her cyst has been misdiagnosed. Of course 31% of the population have bloating of significance [2], but how do we know in this case it isn’t ovarian cancer? So she gets an early ultrasound 3 months later. The cyst has now gone from the ovary, but the other ovary now has a cyst. She gets another scan in 3 months time and the cyst is still there, she’s finds herself more and more worried, why didn’t it go like the last cyst? She reads online about a blood test for ovarian cancer, the CA125. She reads survivors stories telling her the importance of having this blood test done early, so she goes to the doctor and asks to have it done. It comes back slightly elevated. Her fear is confirmed, she has cancer. Now a raised CA-125 has a positive prediction rate of about 10%, and with her imaging findings the likelihood is likely lower, but it is not zero. So we proceed to biopsy. A couple of weeks later the result is in, no cancer, in keeping with the most likely outcome in the scenario. The patient elated at the news thanks the doctor and all is well. Her journey has been 6 months all in all, she’s had multiple sleepless nights, her blood pressure has gone up and her stress levels have been higher, slightly invisibly nudging her up risk of a stroke or other cardiovascular disease in the future.

Now is this good medicine? I guess that’s up for debate, and like I said at the beginning patients like when we scan them, and appreciate when we tell them that their biopsy is negative. They like seeing things done. The doctor who told her not to the scan was clearly a hack as it showed the cyst. Despite the fact that if she’d listened to them she’d have saved herself months of worry and ultimately her health would have probably been slightly better through having avoided the stress and an invasive biopsy. We also know that screening for ovarian cancer does not change mortality for ovarian cancer, it leads to 1% of all women screened having some form or surgery who do not end up having cancer and 3-15% of these women end up with a major complication from this surgery. [6]

Another statistic that I keep in mind is that 11.5% of people under 40 have a thyroid cancer at autopsy and 13.4% of people over 80 [5]. These people lived a good chunk of their lives with this cancer which never caused them any issues or harm, it lay there growing slowly completely undetected and then they died of something else. Now would these people have been better off if they’d got a whole body scan, picked up the cancer and spent the last year of their life having their thyroid gland removed, taking new medication to replace their thyroid hormone, having regular bloods and follow up, all for something that ultimately never would have caused them issues, again I’m not convinced. The patient themselves however if we did go down that route will come in and thank me for saving their life, they’re often so grateful and happy that the cancer was picked up, sometimes they come in with a complication from the surgery, their voice horse from the vocal cord palsy, but they don’t mind as their cancer has been cured. The cancer that would never have caused them any harm.

[1]https://ascopubs.org/doi/abs/10.1200/jco.2008.26.15_suppl.55... [2]https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3264926/ [3]https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6583394/ [4] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7592785/ [5] https://academic.oup.com/jcem/article-abstract/107/10/2945/6... [6] https://pubmed.ncbi.nlm.nih.gov/29450530/

Are you able to provide more information on the fine tuning? Any improvement in WER and what language it was fine tuned in and the size of the dataset used?

2015