The problem was never generating leads. The bottleneck is validated results. It’s performing clinical trials proving efficacy. And this is where I think AI will have the biggest impact—-helping scientist, navigate study complexity and regulatory burden
Drug design is hard because it’s a system problem not a molecule-protein interaction problem.
At one point the gold standard for generating leads was using combinatorial chemistry to experimentally screen through millions of compounds.
I’m not sure how wide-spread that is any more; it used a lot of reagents. And it ran into the real bottleneck: lead generation tells you little to nothing about the efficacy, toxicity, and side effects of a drug or biologic in actual humans.
In silicon will never do a better job at lead generation than combinatorial chemistry. It could help with the real bottleneck. But… there just isn’t a lot of high quality training data on this.
They’ve been pushing ML approaches to drug design since at least the early 2010s. The approach has evolved with the trends in ML. Overall, I think the results have been very mixed; much like the traditional pharma pipeline.
> At one point the gold standard for generating leads was using combinatorial chemistry to experimentally screen through millions of compounds.
I don't think this was ever the gold standard. I mean, some companies did employ these approaches, but they were generally viewed as low-quality, brute-force approaches (at least at the handful of companies I was associated with).
That is not true. For many diseases we lack good hypotheses about the underlying causal mechanisms which resulted in decades and billions being pumped into few questionable targets (e.g. amyloid and alzheimers). In other diseases such as many cancers we often lack "actionable" targets, resulting in hundreds me-too companies all chasing the same target with only subtle variations in approach (inefficient capital alocation).
So AI could help a lot in discovery, much more than the engineering like uses in antibody/drug optimization.
"The problem was never generating leads" is an absurd statement without any reference to the quality of those leads. Ideally you want a system able to generate one more lead, the correct one.
You know what I call solar--fusion power. There's a giant fusion reactor running for free. We simply need to extract the useful energy. Want to build a terrestrial fusion reactor? A conventional nuclear reactor? You still need to extract useful energy and deal with excess heat. The way I see it, from an economic perspective, long-term, solar wins. Of course, this is just my long-view
I installed solar panels in my backyard, and then texted my dad that my basement is now powered by fusion with a roughly eight minute delay.
He then asked "isn't pretty much all power on earth outside of nuclear fission technically done by fusion if we give enough of a delay?". I didn't really have a rebuttal to that.
Geothermal is not nuclear fission. It's some combination of primordial heat from the formation of the Earth (which was gravitational energy) and radioactive decay (which is distinct from fission).
Most of the heavy elements beyond ruthenium or so were created by neutron star mergers rather than red giant or supernova nucleosynthesis[1]. It could be argued that this is technically fission, since neutron stars are, to a first order, giant atomic nuclei.
The lack of a progressive income tax structure gives texas few levers to ameliorate budget issues without cutting services or raising property taxes. It doesn't help the current governor seems to be pushing city budgets to the brink by also cutting off the property tax lever.
Out of curiosity, why couldn't property taxes be progressive? What stops Texas from having different tax rates based on the value of the property?
It even seems like it avoids a lot the pit falls the uber-wealthy use to avoid most progressive income taxes. Its easy to disguise income as something else. Property, though, is right out in the open. The state can easily audit and value property and send the bill. No matter how fancy the accounting, that bill will eventually reach the person "consuming" that property.
In practice, at least from what I've read, in TX, the largest/wealthiest companies get the biggest/best tax breaks (eg Tesla), and end up paying proportionally less, so it's somewhat regressive.
In Texas, like Florida, you get a bill for all improvements within your neighborhood or development as they come. Just because the tax may appear low on the front end, that doesn’t mean you don’t pay for the same relative lifestyle as everybody else in the United States of America.
Even though all the states have their differences they all live relatively the same, in short, if you are an average person, you pay pretty much the same as any place else. It’s only the wealthy who get to hopscotch and dance over the top of all the bodies of the (working class) below.
There’s nothing California can learn from Texas, but if you want to live like Texas or Oklahoma, you can just move to Bakersfield, California hint: that part of California sucks.
And why won’t the frontier models continue to become better? The open models are getting better but so are the frontier models. The frontier models might remain in a constant race to remain ahead
I don't think it's a matter of "staying ahead"; the proprietary frontier models are better, but the trouble for them is that open weight models are good enough in increasingly many cases. This raises the floor on the frontier companies and cuts their total addressable market by commodifying the easier LLM tasks. This is really the central argument of tfa :)
The big question is if there are larger economic gains to be made from ever greater intelligence. It seems to me like that may be the case for only a select few hard problems, while the vast majority of tasks approach their economic ceiling asymptotically with intelligence.
You suggest that the market segment for frontier models is necessarily shrinking. But what if it's just a failure of imagination on the part of users ?
They are running out of novel, clean training data and compute. There is probably a limit to how much improvement can be squeezed out of LLMs. Recent improvements have been more about orchestration and "reasoning" loops (i.e. iteratively feeding context back through the model).
For base models they really must be running out of new training materials. It's more about size and architecture. But they still seem to be getting big strides out of improving the post training. They keep dropping point upgrades in under 8 weeks lately, which is an insane pace for product release. At some point it's going to slow down but we're not close yet imo.
I suspect due to model distillation, their need to stay on top for IPO value is more important than absolutely crushing their competition, who will simply harvest their output and distill it for training data on a ~3 month delay. Also they are probably hitting a wall on increase in intelligence vs training time.
