←
Pradeep Goel4,498 posts
Pradeep Goel avatar

Pradeep Goel

@_PradeepGoel

Founder & CEO 30+ years experience in public health, finance, technology & insurance as CEO, COO, CIO and CTO & warrior for user rights.

◷ Joined October 2017▧ 191 Photos and videos
Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·8h

“If the AI runs on our own hardware, are we compliant?” It’s one of the first questions practice owners ask about private AI. The honest answer: not on its own. Any vendor who tells you otherwise is selling a shortcut. Here’s what running AI locally does solve. Patient data doesn’t have to leave your building for your team to get the benefit of AI. Here’s what it doesn’t solve. Compliance is still a team job: → Your compliance team decides which information the AI can use. → Your clinical leads check the outputs before anyone relies on them. → Your IT team sets the access controls and keeps the logs. What changes is who is in control. With cloud AI, you’re relying on someone else’s promise about your patients’ records. With AI on hardware you own, the rules are yours to set, check and change. If a model, a vendor or a regulation changes, you can adapt without handing your data to anyone. After more than 30 years in healthtech and IT leadership, I think this is the conversation healthcare leaders should be having. Not “Is AI safe?” But “Who holds the keys?” If you lead a clinic, hospital or health system: what’s the first question your team would ask before approving AI on patient data? #PrivateAI

Post attachment
Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·13h

Thalamus Decision is not a replacement for laboratory validation or scientific judgement. Its potential value is in helping researchers prioritize experiments, narrow down candidates and make better use of limited research resources. The next important test will be whether these results hold up across independent datasets and real drug discovery workflows. If they do, typed decision models could become a valuable part of how biological research moves from data analysis to experimental decisions.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·13h

These figures deserve attention, but they also need careful interpretation. A target error rate is not the same as the error rate actually achieved, and benchmark performance does not establish how reliably a model will perform in clinical or experimental settings. I find the decision making approach especially relevant to biological research. AI systems that can communicate uncertainty provide a more useful basis for human oversight than systems that simply produce confident-sounding answers. Researchers need to know not only what a model predicts, but also when its prediction should be treated with caution.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·13h

Some of its reported benchmark results are notable. It achieved 84–89% on two drug-response prediction tasks and approximately 94% on gene essentiality prediction across tissues. The confidence based results are particularly interesting. At a target error rate of 5%, the 9-billion-parameter model automated 62.6% of decisions, while the 27-billion-parameter version automated 66.2%. Their reported realized error rates were 6.3% and 6.8%, respectively.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·13h

Unlike a conventional language model that generates a written response, Thalamus Decision assigns confidence scores to possible outcomes. This allows a system to distinguish between decisions it is sufficiently confident to automate and cases that should be referred to a human. The model can analyze bulk and single cell RNA sequencing data for tasks such as identifying cell types, assessing malignancy, determining tissue of origin and predicting how cancer cell lines may respond to drugs. It can also predict how silencing a particular gene may affect gene expression.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·13h

One of the challenges in drug discovery is deciding what deserves to move from computational research into the laboratory. Researchers can identify promising molecules, study biological pathways and generate large volumes of data, but each potential candidate still needs to be evaluated before significant resources are committed to experimental validation. Ingenix, is approaching this problem with Thalamus Decision, a model designed to make structured decisions directly from biological data.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·Nov 17

The industry will need to address the boundaries around model outputs and intellectual property without treating every form of knowledge transfer as misconduct. That distinction will matter as distillation becomes a more established part of AI development.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·Nov 17

What stands out to me is how the economics of AI development are changing. Restrictions on advanced chips can make it harder to train and operate large models, but they do not prevent developers from finding other ways to acquire useful capabilities and deploy them more efficiently. This is why the AI race cannot be understood through hardware restrictions alone. Research, engineering efficiency, access to knowledge and the ability to turn existing capabilities into practical products are becoming just as important.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·Nov 17

There is an important distinction here. Distillation is a recognized AI development technique, and its use is not inherently illegal. However, collecting another company's model outputs at industrial scale to train a competing product raises legitimate questions about intellectual property and the terms under which those outputs can be used. I would also be cautious about the suggestion that distillation alone explains the narrowing gap between Chinese and American AI models. The researchers and industry sources cited in the report point to other contributions, including improvements in algorithms, engineering, data and model design.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·Nov 17

The approach is particularly useful when access to advanced computing hardware is restricted. Instead of spending enormous resources trying to develop every capability independently, developers can use outputs from an established model to build selected capabilities into a cheaper system. The investigation documents how this is moving beyond isolated research experiments. Chinese companies have developed processes to generate, refine and score training material from American models, while government policy is also encouraging distillation for industrial applications.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·Nov 17

China's growing use of American AI models to train domestic systems shows how difficult it is to contain the spread of AI capabilities through chip restrictions alone. A Reuters investigation found that Chinese researchers and companies are using outputs from models such as OpenAI's GPT-4 and o1 to develop training datasets for smaller domestic models. This practice, known as knowledge distillation, allows a less capable model to learn from the answers, reasoning examples and task demonstrations produced by a more advanced one.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·Nov 16

In aviation, efficiency cannot come at the expense of safety. The measure of SMART's success will be if it consistently improves air traffic management while giving controllers reliable information they can trust.

Pradeep Goel avatar
Pradeep Goel@_PradeepGoel·Nov 16

One distinction is important here, SMART currently recommends actions; human specialists and air traffic controllers remain responsible for decisions. AI handles the computation, but it does not have authority over flight operations. I believe this is the right direction for deploying AI in safety critical environments. These systems can help people process information and identify options that might otherwise be missed. But recommendations must be verifiable, and experienced professionals must retain the authority to reject them.