kevinalexbrown

"tell me about biology" -> "Switched to Opus 4.8"

Building the fndamental human patient representation: https://standardmodelbio.substack.com/p/the-patient-is-not-a...

The patient is not a document - multimodal foundation models for biomedicine. JEPA's working well.

Standard Model Bio | various | Full-Time | ONSITE SF or Philadelphia | Hybrid

we're building the standard model for bio. We're doing for biology what mathematics did for physics.

papers this month: genomics - https://arxiv.org/abs/2509.25573 protein-language: https://arxiv.org/abs/2509.22853 longitudinal ehr: https://arxiv.org/abs/2509.25591

hello world: https://standardmodelbio.substack.com/p/introducing-standard...

we're humble and ambitious - we want to be the quiet backbone of biomedical ai but we want none of the glory of the final applications.

kevin@standardmodel.bio

building the standard model for bio: https://standardmodelbio.substack.com/p/introducing-standard...

We're pretty jazzed.

Yes and this response has also become a trope. It’s generically universal to nearly any early result.

The QZ article is narrowly correct but widely misleading. It almost willfully ignores the momentum and direction.

In reality, radiologists will not be summarily replaced one day. They will get more and more productive as tools extend their reach. This can occur even as the number of radiologists increases.

Here's a recent example where Hinton was right in concept: recent AI work for lung cancer detection made radiologists perform better in an FDA 510k clearance.

20 readers reviewed all of 232 cases using both a second-reader as well as a concurrent first reader workflows. Following the read according to both workflows, five expert radiologists reviewed all consolidated marks. The reference standard was based on reader majority (three out of five) followed by expert adjudication, as needed. As a result of the study’s truthing process, 143 cases were identified as including at least one true nodule and 89 with no true nodules. All endpoints of the analyses were satisfactorily met. These analyses demonstrated that all readers showed a significant improvement for the detection of pulmonary nodules (solid, part-solid and ground glass) with both reading workflows.

https://www.accessdata.fda.gov/cdrh_docs/pdf20/K203258.pdf

(I am proud to have worked with others on versions of the above, but do not speak for them or the approval, etc)

The AI revolution in medicine is here. That is not in dispute by most clinicians in training now, nor, from all signs, by the FDA. Not everyone is making use of it yet, and not all of it is perfect (as with radiologists - just try to get a clean training set). But the idea that machine learning/ai is overpromising is like criticizing Steve Jobs in 2008 for overpromising the iphone by saying it hasn't totally changed your life yet. Ok.

Siemens Healthineers Imaging Intelligence | https://www.siemens-healthineers.com | Malvern, PA (Greater Philadelphia) | INTERNS | Onsite Our R&D group delivers medical image/text tools (e.g. deep learning, NLP, etc) for medical data analysis. We are well recognized for delivering cutting-edge intelligent solutions to Siemens 3D workstations and medical imaging scanners. Our group also has strong publication record in top tier journals and conferences, and several Siemens "inventor of the year" award recipients. We offer well-paid internships lasting >= 3 months, with independent moonshot projects.

Responsibilities: · Contribute to research projects to develop intelligent solutions for medical imaging and text analytics · Conduct fast prototyping, feasibility studies for exploratory clinical research · Support the productization of research prototypes

We look for: · Strong research capability in computer vision, machine learning, text analytics and medical image analysis, proven by publications in journals/conferences. · Research experience in image/text analytics using large scale, weakly supervised / unsupervised learning algorithms · Research experience in medical image/text analysis of different modalities (CT, MRI, PET, medical reports etc.)

Email: Kevin.Brown@siemens-healthineers.com

Siemens Healthineers Imaging Intelligence | https://www.siemens-healthineers.com | Malvern, PA (Greater Philadelphia) | INTERNS | Onsite Our R&D group delivers medical image/text tools (e.g. deep learning, NLP, etc) for medical data analysis. We are well recognized for delivering cutting-edge intelligent solutions to Siemens 3D workstations and medical imaging scanners. Our group also has strong publication record in top tier journals and conferences, and several Siemens "inventor of the year" award recipients.

We offer well-paid internships lasting >= 3 months, with independent moonshot projects.

Responsibilities: · Contribute to research projects to develop intelligent solutions for medical imaging and text analytics · Conduct fast prototyping, feasibility studies for exploratory clinical research · Support the productization of research prototypes

We look for: · Strong research capability in computer vision, machine learning, text analytics and medical image analysis, proven by publications in journals/conferences. · Research experience in image/text analytics using large scale, weakly supervised / unsupervised learning algorithms · Research experience in medical image/text analysis of different modalities (CT, MRI, PET, medical reports etc.)

Email: Kevin.Brown@siemens-healthineers.com

If you find this kind work interesting, our AI group at Siemens Healthineers is hiring interns to carry out projects like this. We typically target machine learning or medical imaging PhD students, but are open to a variety of backgrounds. Please feel free to reach out via email.

Siemens Healthineers Imaging Intelligence | https://www.siemens-healthineers.com | Malvern, PA (Greater Philadelphia) | INTERNS | Onsite

Our R&D group delivers medical image/text tools (e.g. deep learning, NLP, etc) for medical data analysis. We are well recognized for delivering cutting-edge intelligent solutions to Siemens 3D workstations and medical imaging scanners. Our group also has strong publication record in top tier journals and conferences, and several Siemens "inventor of the year" award recipients.

We offer well-paid internships lasting >= 3 months, with independent moonshot projects.

Responsibilities: · Contribute to research projects to develop intelligent solutions for medical imaging and text analytics · Conduct fast prototyping, feasibility studies for exploratory clinical research · Support the productization of research prototypes

We look for: · Strong research capability in computer vision, machine learning, text analytics and medical image analysis, proven by publications in journals/conferences. · Research experience in image/text analytics using large scale, weakly supervised / unsupervised learning algorithms · Research experience in medical image/text analysis of different modalities (CT, MRI, PET, medical reports etc.)

Email: Kevin.Brown@siemens-healthineers.com

I read the original study. I have a few thoughts. My partner is a physician and I am an AI researcher working in medicine. I think a lot about doctors as machine learning models, and RCT results like loss terms in a complicated objective function.

What is the best learning rate for updating physicians (our models) from the results of RCT's (part of our loss)?

The authors reviewed all articles in three journals from (generally) between 2003-2017. They didn't, afaict (please point me if they did), review the time-to-correction (if any correction has been made). It takes some time before the results of an RCT end up in established practice. I'm actually surprised it's so small in many cases.

It's not like there's a database where the results of every RCT are immediately updated and the physician model is retrained overnight on the new data.

Even if there were, imagine if the learning rate (so to speak) were so high that every discipline immediately changed their published best practices on the basis of a single RCT?

Here's a cautionary paragraph from one of the excellent reversal studies they use:

Several limitations of the study warrant discussion. First, because we enrolled only 26% of eligible patients, our findings must be generalized cautiously. The most frequent reason that patients declined enrollment was a strong preference for one treatment or the other. Since patients' preferences may be associated with treatment outcome, our trial may be vulnerable to selection bias. Participating surgeons may not have referred potentially eligible patients because they were uncomfortable randomly assigning these patients to treatment; this form of selective enrollment may also create bias.26 Second, because the trial was conducted in academic referral centers, the findings should be generalized carefully to community settings. Third, we did not formally assess the fidelity of the physical therapists or surgeons to the standard intervention protocols. Finally, our study was not blinded, since our investigative group did not consider a sham comparison group feasible. [0]

I'm less concerned about RCT to Best Practice time than from Best Practice to Typical Physician Practice time. There is a cascaded model connected to the 'complex RCT loss' and it's discipline published practice down to individual physician treating patients. Compressing the time from RCT to individual physician is fraught with difficulties, but could be improved.

Finally, RCT is the gold standard, but it's not perfect and it doesn't always clearly translate to the individual physician's model of practice. Many best practices weren't established from RCT's either.

And an inconclusive result from an RCT is not the same thing as proving that there's no difference in outcomes, but a proper statistician can chime in there.

[0]https://www.nejm.org/doi/full/10.1056/NEJMoa1301408

It's funny, I was thinking that Inbox did this nicely, and of course he did that too.

Some medical devices also have other net cost benefits though. For example, shorter, less intensive hospital stays from less invasive testing.

Dr Khullar suggests that AI will exacerbate biases in medical practice. His fundamental concern is that machine learning will codify biases and become self-fulfilling prophesies. But there is scant evidence that AI will worsen these disparities.

If anything, a machine-learning point of view better addresses his concerns than a traditional one, because they can be much more quickly updated to correct for identified biases. Doctors spend years and years of hard work becoming efficient and effective human algorithms themselves, and updating those human algorithms in the face of newer evidence is difficult. In standard practice, biases are often invisible and uncodified to begin with. "Moral intuition" is something all doctors use, but it's also something of a black box in nearly every real-world use case.

His writing is somewhat tongue in cheek.

For whatever it’s worth, he seems to be a dedicated teacher who posts self criticisms of his courses publicly online. The course this book is based on has grown quite successful as well.

I’m not sure if you’re just trolling, but encouraging others to avoid this book might be a mistake. It’s quite good.

The packages-first attitude feels significant to me. A language that “users” enjoy but package developers also enjoy seems important. I hadn’t thought about language choice from a heavily package-development weighted perspective before. It seems obvious in retrospect though, which is probably a sign of something cool.

A version of this would be: how can good package development be as easy as possible, and how can package use be as easy as possible?

I haven’t done any serious work in Julia mainly because the python libraries are mature, good, and performant enough. I can’t speak for everyone, but for end users in science labs library support is perhaps the biggest consideration for language choice.

This is an important point, but one I feel the article acknowledges well.

There are many impossibility problems that can be solved by relaxing some constraints (like wait-free) or by accepting some unsolvability (we don’t worry too hard about our programs halting in practice, and we generally trust compressed sensing results because the probability of failure is provably delta, say), or just by changing some desiderata.

Great examples include arrows impossibility theorem or the no good clustering result, each of which have solutions if the assumptions that operationalize our intuition are changed slightly.

Because the article just offers a single example, and seems to center on the lack of available tenure track positions, I'm going to take a more holistic view that gets at the major issue: is current research funding at the appropriate level?

What would happen if we increased research funding by X percent? How did we settle on the current funding levels? I would be curious to see a reasonable source for this. A cursory google search mostly returned opinion pieces that we should increase funding for science. I agree, but hard(er) numbers would be better. It would be great to see a back-of-the-envelope ROI for X percent funding increase in T time. Obviously funding can be applied in many ways, and the ROI is difficult to measure, but someone must have studied it.

For the immediate future, the US remains the best place for research. But dominance can begin to change before the effects become obvious, like a large company that's still profitable long after it's become irrelevant.

I'm more curious about the byproducts and energy consumption of each manufacturing process. I have no idea, does anyone have a good comparison?

Aside: supposedly global pulp production is 34 percent recycling, 45 percent from sawmill waste, and 21 percent 'logs and chips.'[0] The wikipedia article later states, from another source, that 16 percent of production comes from tree farms. The gist of your general point stands, as I understand it: we're not going out and cutting old growth or even secondary growth forests for paper (though we make use of reject trees when we target them for other reasons).

[0] https://en.wikipedia.org/wiki/Pulp_(paper)

I remember once from undergrad the saying '"obvious" is the most dangerous word in mathematics.' Seeing how something could be true is dramatically different from identifying and defending that it is true.

It's dangerously easy to say "oh yea, makes sense, natural selection happens by mating so if mates choose club wings, I get it. Obvious." But Prum's trying to go a step further, and test just how far out of balance and arbitrary the mate selection part can be from the direct do-not-die part of evolutionary fitness.

He proposes that we can differentiate between these two by considering that the club wings aren't actually indicators of higher direct fitness, because they hurt the ability to fly, even among females that have no need for such shenanagins. I'm not sure I totally agree with or grasp that, but at least it's an attempt to further understand and test the idea.

I'm frankly surprised by comments accusing a well established evolutionary biologist of severely misunderstanding natural selection. The author has spent his career, among other things, investigating mechanisms of evolution, and identifying and performing tests to assess their relative importance to a particular species (here's an example: http://prumlab.yale.edu/sites/default/files/prum_1997_phylog...).

You might consider whether your objections are addressed in his work not aimed at the lay population, and that your criticism really just amounts to "He wrote this at not exactly the right level of sophistication for me." Maybe that's true, but it's a pretty boring claim.

But the evidence supports the idea that social isolation induces measurable changes in behavior and neural activity.

And you can definitely make causal inferences if you don't understand everything. If that weren't true, you wouldn't be able to infer that moving your arm makes the coffee cup in your hand move unless you knew everything about physics.

I'm not sure that's true here.

They went further than showing a correlation between two traits. They directly manipulated the pre-alcohol state (alone, or with other crayfish), and measured whether that change outcomes in behavior and neural activity. Presumably they kept everything else the same[0], randomly assigned crayfish to isolation and social environments, and placed them in the same alcohol environment.

Just because the chain of causation is fuzzy doesn't mean A doesn't cause B. Maybe social isolation and alcohol response are trivially connected to something else like movement rate and other variables. That doesn't mean you can't make causal inferences.

[0] I don't know much about crayfish experiments - maybe there's some latent cause like isolated crayfish are handled differently and that causes changes in alcohol-induced behavior and neural activity. I would want to know whether the crayfish showed similar or divergent behavior in an alcohol-free environment. It's possible that alcohol has identical 'gain' on behavior and neural activity, but that still indicates an effect, technically. And nothing is ever 100 percent the same across conditions (this is the fundamental problem with causal analysis) but that's true across every discipline.

If there's an objection to the title, it's that loneliness and social isolation are not the same thing.

A friend of mine from academia was considering going into industry, so went to some data science meetups. Someone was giving a presentation about convolutional networks, yet did not know what a convolution was. At first I was startled, but in the long run machine learning applications will be decided just as much on user experience and design features as on algorithmic choices.

I'm not a web programmer, but I imagine few developers could remember the mathematics[0] of the sorting algorithms that are fundamental underpinnings of their work (if they ever learned them at all). Yet I'm not sure it matters, even to great developers. The same thing will probably ultimately be true of machine learning. Honestly, you need not know what a convolution is to build a perfectly usable convnet. (And ultimately you may not need to even build your own if you can use a nifty amazon API.)

Whether NIPS should care or not is a separate story. It seems a little sad - I took all this hard pure math as an undergrad, and it doesn't seem to be important if all I'm doing is changing a few layer parameters (even if the change is ingenious).

[0] mathematics as in proof of sorting, proof of bounds on space/time complexity, etc.

Hi Eric, I haven't read the paper in depth yet, but we've discussed it some as a lab. I love the concept, but I'm a little perplexed by something I hadn't thought of when we first discussed it.

It looks like many analyses were performed by binning transistor switching as you might spikes from neurons, or by linearly combining transistor activity across transistors or time.

This is slightly puzzling to me, because transistors would appear to not work on the basis of average activity at their level of computational composition. I would be surprised if you could really understand circuit activity at a computational level with most of the measurements you describe.

On the other hand, I suspect the methods used work slightly better in a real brain. Of course, the computational purpose of many neurons is difficult to parse or even construct useful hypotheses about. But at the very least, neurons in the outer periphery seem to behave in a way that is at least consistent with firing rate, e.g. impulses at the neuromuscular junction to cause muscle contraction, and retinal ganglion cells to fire at higher rates when the input matches their certain luminance features.

I don't think anyone really claims that all you need is the wiring diagram, even those who really want to see a connectome. Then again, I stay away from philosophical debates about the connectome.

Everyone who's paying attention understands that individual neurons have the potential for very complex, but precise, behavior.

But knowing the connections across the brain (at some resolution) is helpful for plenty of reasons. If I want to understand how areas in the brain communicate, it's immensely useful to know where they're connected, for instance. Let's say I have 200 sensors I can place in the brain wherever I want. Placing them at crucial nodes or connected areas could be tremendously useful, since we can't yet put sensors everywhere for most spatial and temporal resolutions we want.

If you live in Seattle, check out the Mountaineers. If in Portland, the Mazamas. First and second largest mountaineering groups in the US iirc. I learned to climb through maxamas, you'll learn to do it right. They don't just climb though, lots of hikes and training.

Were you seen in a hospital or a private practice? If it was a hospital you almost certainly primarily interacted with graduates of the last 5 years since you're mostly seeing residents.

Fwiw, "putting in effort" is somewhat relative. Do you mean they were lazy, or spread too thin?

Without knowing this specific case, I can't say for sure. But my broader point is that the knowledge base required of doctors is so huge that even if it's not excusable, it is understandable. Like online community moderation, it's really easy to point to cases where doctors got it wrong and assume it's all for naught.

Absolutely the accuracy of doctors should be improved, but the bar is not "are you as good as a UCSF resident?" (one of the best programs nationally) but "are you better than nothing?"

It just seems crazy to me to suggest that the median doctor, i.e. at least half, is worthless.

the median doctor is pretty worthless

This seems fairly disrespectful. I suppose you mean that the median doctor isn't perfect, but do you have any idea how many hours residents put in? You might think carefully before calling the median one worthless, and ask whether they're really, truly wasting their lives in the service of others.

Less emotionally, I find it hard to believe that half of doctors proved essentially no benefit.

it's a side effect that's been widely known since the 1970's

Guess how many side effects have been widely known since the 1970's?

For clarity: I am 100% in favor of thinking critically about a doctor's advice, having received contradictory diagnoses before.

It was an image segmentation task, and the features were similar across data sets. The other thing that made it work well was heavy use of data augmentation that captured ways in which different data points could reasonably differ.

There was a really cool medical imaging paper recently that literally just labeled several 2D slices in a 3D dataset consisting of 3 images and performed a reasonable segmentation:

https://arxiv.org/abs/1606.06650