2024 Big Ideas: Miracle Drugs, Programmable Medicine, and AI Interpretability

26:53Or where does it have the potential to really reshape that arc? Yeah. So, I think this is where the concept of a programmable medicine potentially could be very transformative to how we think about developing drugs. And that is because in the programmable medicine, there are multiple components that could be, as I described, redeployed for different applications. So let me give you an example. Aging therapy is this idea that you can deliver a genetic payload. In other words, you can deliver a gene to a cell that has a defective version of that gene. If we're able to make a medicine that does that, that can deliver one gene to a given cell type, it becomes increasingly likely that we'll be able to deliver a different gene to a different cell type for a different disease.

27:40And that's very different than the way we traditionally make medicines. If you think about a traditional medicine, like a chemical, a molecule, that is a chemical that is tested and designed for a specific disease, a specific target. The second you switch the target, we don't reuse the atoms in the molecule and try to fit them into a new target. We just design a different molecule. That's why I say in traditional drug development, one molecule has very little bearing on the next molecule you design. But in the case of these programmable medicines where all the components can be reused, you You just have to essentially redirect or redeploy components like delivery vehicles for gene therapies or in the case of a gene editing medicine, just redirect what edit you want the enzyme or the protein to make.

28:33And that is where the programmability comes in. And that sounds huge, right? Just so I understand you correctly, when you're talking about the reusability of these medicines, It's basically the equivalent of the reasonable rocket is almost like you basically would have already gotten a certain drug approved to be able to deliver something each time you're reiterating on that. Would you then need to have the specific genetic almost like payload in there, re -approved, but most of that like work has already been done. Am I thinking about that correctly? That's right. In this analogy, what we would love to see become possible is that the rocket, in this case, is the vehicle by which you ensure that your payload is delivered.

29:16Right? And some of the most common ones that we know in terms of what would be rockets here are the LNP molecules that all of the COVID vaccines that many of us received were delivered. The COVID vaccine was mRNA. mRNA had the instruction for what it wanted your body to make. And that was encapsulated in this LNP particle, lipid nanoparticle. It's like a little ball of fat. Similarly, for a lot of gene therapies, instead of using an LNP particle, we use something called an AAV, which is an adapted virus. It stands for a Deno -associated virus. There have been examples in the clinic where palids have been delivered with LNP or palids have been delivered with AAV.

29:59For different applications, all you're swapping out is the cargo, the instruction of mRNA that's in that LMP or the instructions or the genetic cargo that's in that AAV. And this is very timely because the FDA is starting to signal that they are looking for ways to be increasingly adaptive to ensure that they can adequately review these therapies that are reusing components, but do so in a way that will be both rigorous but also adaptive and therefore hopefully more speedy. And where people don't have to start from scratch. And you don't have to start from scratch. Because I already have that rocket that is replicable.

30:37So tell us a little more about that, where the FDA seems to be taking maybe some inspiration or at least that's what you alluded with the FAI, right? So what can they learn from the rocket reusability in this analogy? And what also may be fundamentally different? And what are you taking from some of the new announcements like the new office of therapeutic products? So I think there's a lot of things to take from this. The first one is that I think there is real recognition here that we are seeing meaningful innovation in terms of the kinds of medicines we can make. And the FDA is appropriately trying to find ways to retain their rigor because again, they have an extraordinarily high obligation to ensure safety in efficacy in patients.

31:19but to move hopefully more quickly to attempt to keep up with all the innovation we're seeing. And so they pushed forward on several fronts that I think are already pointing us in that direction. The first one is they've announced an office of therapeutic products whose mandate and mission is to find ways to do exactly what we've been describing. And so I think the industry, the drug development industry now has a partner in the FDA in trying to find the best paths forward along the lines of what we talked about. The second is they are also trying to innovate themselves. The FDA is looking to be more nimble and is running experiments to find ways to do exactly that.

32:06So for example many of us would be familiar with Operation WorkStreet which was the effort by the government to try to ensure that the COVID vaccines could reach all of us in a very, very timely manner, given the nature of the pandemic. Well, the FDA has said there are so many intractable diseases, many rare genetic diseases that have no good therapeutic or treatment options. So they are launching a pilot program marrying the concept of operation warp speed, but to apply that to rare diseases, where they are going to run some experiments to see different ways that the agency can interact with industry to get these medicines more quickly to the patients that so desperately need them.

32:49And so I think those are very, very promising signs that both the industry and the government are looking for ways to ensure these innovations reach patients in a timely and responsible manner. And there's also other proof points that we can point to that are happening every day right now. So, for example, several companies that are developing cutting edge gene editing medicines are starting to get approvals by the FDA to move forward with some of their key clinical trials. And there had been a moment in time where the FDA was being more hesitant, but I think as they've started to evaluate these technologies more carefully, they've started to develop path forward for these medicines to continue to get developed.

33:28And the breaking news is that the first CRISPR therapy has been approved. This is a therapy for sickle selenemia and beta thalassemia that was developed by a pharmaceutical company called Vertex Pharmaceuticals, working with another company called CRISPR Therapeutics. And this is a therapy for essentially curing a genetic disease, in this case, sickle selenemia, by taking the relevant cells out of the body, editing them with CRISPR gene editing technology, and putting them back in the body in a way that results in a functional cure. And this is a big deal for so many reasons. The first one, it is the first time that a CRISPR therapy has been approved as a medicine.

34:09The second reason why this is a big deal is that it gets an important milestone to this point about the future of what programmable medicines could look like. Now that you have a first approval, you've demonstrated that a CRISPR editor could be safely used as a medicine. So now that it adds something differently, the process for getting that approved should hopefully be shorter and faster. And the third reason I think that this is a big deal in terms of a milestone is relatively speaking how quickly this has happened. So you asked me at the beginning, how long does it take to make a drug? And I said about 10 to 15 years on average, It's barely been 10 years since the concept of CRISPR was discovered and described in the scientific literature.

34:52Right. So we went from the initial discovery of CRISPR as a potential use, as a medicine, all the way to an approval in just over 10 years. That is lightning fast in this world. So it's just an exciting moment in time, I think, for the industry and hopefully for all the various patients that are looking for better solutions, the treatments for their disease. Absolutely. And I mean, it really does feel like a new paradigm or at least we're moving towards that where you see this combination of changes in innovation and regulation coming together where you see things like the news that you just shared.

35:25Maybe you could just speak a little more to that like if we are able to see this confidence that you're discussing and this idea of programmable medicine becoming a reality. It was not really mean in terms of like the speed of therapies, these coming, the number of them, the new business models that might be unlocked, how this might ultimately end up impacting patients. Any just kind of high level thoughts about if we are moving to this new paradigm, what that really means. Yeah, I think it means a couple of things. The first one is, I think it means that for the benefit of patients, every time we run a clinical trial with these types of medicines, we're not starting from square one, because we already know something about the various components in the therapy.

36:07So that's the first thing. The second thing is that generally speaking, these programmable medicines are going after diseases where the cause of the disease is very well known. In other words, it's a known mutation and it's the ability to intervene in an effective way has been what's been a luciferous. So in a lot of ways these new programmable medicines are just a fundamentally new superpower. We can go after diseases that we weren't able to go after before. And as a result, the third thing that I think this means for all of us is that we may be on the cost where the elusive seaworth is a reality that we might actually have cures for lots of very intractable problems.

36:49And that is a very new day indeed, right? Like that is something that it just has not been very common in our industry. So I think there's lots of reasons to be excited. Yeah, I mean, as you're talking, I'm like actually smiling because it's impossible not to be excited. But I guess just to close things off, you've painted this beautiful picture of what may be to come. Could you share a little bit more about the blockers if any, whether regulatory, whether it has to do with the fundamental science that's coming on board? What would you say if this reality that you're painting weren't to not come to pass?

37:22What would those reasons be? And also maybe for those listening, how can builders get involved? How can they actually help make this reality come true? And first of all, in terms of the blockers, I think there are several, and I think they are important, and some of the blockers should be there. So the first one is everything I'm describing in terms of these programmable medicines, it has another side of the coin to it, which is these are largely speaking permanent medicines as well. So if you take a pill and you have a bad reaction or adverse event or toxic reaction to that pill, you just stop taking the pill.

37:57And eventually your body will clear it and hopefully the toxicity has been addressed. In the case of making a permanent edit in DNA, if there is an error, if there is a toxicity that comes from making that edit something inadvertently, it's permanent or at least it has the potential to be permanent. And so for that reason, the FDA appropriately has extraordinarily high bar for how they think about evaluating the safety of these medicines and how they think about which diseases are probably the most appropriate to go after with something that is potentially a cure, but also potentially permanent.

38:36Yes. So that's one blocker. It's going to need to be addressed systematically. And it should. A second blocker, the only medicines that work are the ones that get to patients. And these programmable medicines have a couple of challenges in terms of accessibility. The first one is these are not pills you get over the counter. These are very complex medicines. And so therefore the process by which you get treated can be a very long and difficult process. So I just described the approval of CRISPR therapy for sickle cell anemia. Yeah. The process by which you get the treatment could take months because you have to go, you have to get your cells removed, the cells get edited.

39:25In this case, you actually have to get a form of therapy to kill your internal cells so you can replace it with the corrected ones afterwards. And that could be a long stay at the hospital. So that whole process could take several months. So the accessibility of these therapies will limit how many people can get them and when. And the second element to accessibility is cost. Yes. And today, these therapies can cost on the order of millions of dollars. Now on the one hand, that millions of dollars of cost is justified because a lot of R &D went into it. They're very expensive to make in manufacturers.

40:05It's very expensive to manufacture all these components I'm describing. And that's number two. And the number three, they provide a fair amount of benefit. In the case of treating a baby that had something like somatic muscular atrophy that would have otherwise died, a one -and -done gene therapy from the artist. This is a therapy that costs about $2 .1 million. Effectively saves that baby's life. So there's an incredible amount of value that comes from that. Absolutely. But the cost of discovering, developing, manufacturing these therapies and the benefit that they come from also country that very hefty.

40:38Price tagging, that's going to limit accessibility. So that's one of the other key blockers, I would say, that we're seeing in this space. Now, how can builders help? I think builders can help in a very important way, which is if there is one technology on this planet, that can scale better than anything else we know of, its biology. And so all of the things I'm describing in terms of blockers to access, at some point can be addressed by improving our ability to engineer biology to address some of these limitations. So improved biology can make manufacturing much more scalable and therefore much less expensive, improving the kinds of interventions and the precision by which these programable medicines work can address some of these questions of permanence or toxicity.

41:28So where the builders can really help is just become better programmer versus biology and from that we will get better applications at higher scale and at lower cost. And hopefully get them in the hands of patients more quickly. A lot of our listeners are familiar with the idea of exponentially decreasing cost and things like software. Do you see that same kind of curve being applied here when you're talking about decreasing cost? Is that really the future that you're painting where these things instead of costing millions of dollars we're talking thousands, is that really in the future? That's the hope.

42:01And the possibility is there. Because again, biology can scale exponentially. We all did come from one cell and here we are. We're sitting here. That's such a great point. Yeah, but there's work to be done there. We haven't seen it yet. But that's the promise and that's the hope. So as our health system looks to the skies for inspiration, what inspiration can we take from our own biology to understand how large language models work? and will we ever move from black box to clear box? My name is Anjane Mita, I'm a general partner here at A16Z and I'm talking to you today about AI interpretability, which is just a complex way of saying reverse engineering AI models.

42:42Over the last few years, AI has been dominated by scaling, which is a quest to see what was possible if you threw a ton of compute and data at training these large models. But now, as these models begin to be deployed in real world situations, the big question on everyone's mind is why? Why do these models say the things they do? Why do some prompts produce better results than others? And perhaps most importantly, how do we control them? Ajay, I feel like most people don't need convincing that this is a worthwhile endeavor for us to understand these models a little better. But maybe you could share where we're at in that journey.

43:18What do we or don't we understand about these L and black boxes and their interpretability? It might help the reason by analogy here because this is a set of abstract ideas, but to make it a little bit more concrete pretend one of these AI models is like a big kitchen with hundreds of cooks. And when you ask the kitchen to make something, each cook knows how to make certain foods. And when you give the kitchen ingredients and you say, take a cook meal, all the different cooks debate about what to make and eventually they come to an agreement on a meal to prepare based on these ingredients. Now the problem is where we are in the industry right now is that from the outside we can't really see what's happening in these kitchens.

43:59So you have no idea how they made that decision on the meal. You just get the cake. You just get the cake or whatever it might be. Right. So if you ask the kitchen, hey, why did you choose to make lasagna? It's really hard to get a straight answer because the individual cooks don't actually represent a clear concept like a dish or cuisine. And so the big idea here is what if you could train a team of head chefs to oversee these groups of cooks and each head chef would specialize in one cuisine. So you'd have the Italian head chef who controls all the pasta and pizza cooks. And then you have the baking head chef in charge of cakes and pies.

44:33And now when you ask why lasagna the Italian head chef raises his hand and says, I instructed the cooks to make a hearty Italian meal and these head chefs represent clear interpretable concepts inside the neural network. And so this breakthrough is like finally understanding all the cooks in that messy kitchen by training these head chefs to organize them into tidy sort of cuisine categories. And we can't control every individual cook, but now we can get insights into the bigger or more meaningful decisions that determine what meal the AI chooses to make. Does that make sense? It does, but are you saying that we do actually have a sense now of those like head shafts or the people responsible for parts of what might be happening within the AI?

45:16Obviously it's not people in this case, but have we actually unlocked some of that information with some of the new releases or new papers that have come out? We have. And you can break the world of interpretability down into a pre 2023 and a post 2023 world in my opinion because there's been such a massive breakthrough in that specific domain of understanding which cooks doing what? More specifically, what's happening is that these models are made up of neurons, right? A neuron refers to an individual node in the neural network, and it's just a single computational unit. And historically, the industry sort of tried to analyze and interpret and explain these models by trying to understand what each neuron was doing, what each cook was doing in that situation.

45:58Of each on the other hand, which is this new atomic unit that the industry is proposing now as an alternative to Neuron refers to a specific pattern of activations across multiple neurons. Okay. And so while a single neuron might activate in all kinds of unrelated contexts, like whether you're asking for lasagna or you're asking for a pastry, a feature which is this new atomic unit represents a specific concept that consistently activates a particular set of neurons. And so to explain the difference using the cooking analogy, a neuron is like an individual cook in the kitchen, each one knows how to make certain dishes, but doesn't represent a clear concept.

46:35A feature would be like a cuisine specialty controlled by a head chef. So for example, the Italian cuisine feature is active whenever the Italian head chef and all the cooks they oversee are working on an Italian dish. And that feature has a consistent interpretation, which in this case is Italian food, while individual cooks do not. And so in summary, these neurons are individual computational units that don't map neatly to concepts. These features are patterns of activations across multiple neurons that do represent clear, interpretable concepts. And so the breakthrough here was that now we've learned how to decompose a neural network into these interpretable features when previous approaches focused on interpreting single neurons.

47:16And so the short answer is yes, we have a massive breakthrough. We actually now know how to trace what was happening in the kitchen when a dish was being made. And maybe could you give an example that's specific to these LLMs when we're talking about a feature? I know there's still so much research to be done, but what's an example of a feature that you actually see represented in the outputs from an LLM? Yeah, that's a great question. So I think if you actually look at the paper that moved the industry forward a bunch earlier this year, there's a paper called Decomposing Language Models with Dictionary Learning.

47:48This came out of Anthropic. Interpreter abilities a large field, but this paper, I think, took a specific approach called mechanistic interpretability. And the paper has a number of examples of features that they discovered in a very small, almost toy -like model, because smaller models proved to be very useful, feed -tree dishes for these experiments. And I think an example of one of these features was a God feature, where when you talk to the model about religious concepts, then a specific God feature that mapped to this concept of a god fired over and over again. And they found when they talked to the model about a different type of concept like biology or DNA or I think biological concepts was one of the sort of types of questions they were asking the model a different feature that was unrelated to the god feature fired.

48:35Whereas the same neurons were firing for both those concepts. And so the feature level analysis allowed them to decompose and break apart the idea or the concept of religion from biology, which is something that wasn't possible to tease apart in the neuron world. Yeah, and maybe you could speak a little bit more to why this is helpful. I mean, maybe it's obvious for folks listening, but now that we have these concepts that we see and maybe can also link pretty intuitively, like, oh, okay, I understand biology. Oh, I understand religion as a concept that's coming out of these LLMs. Now that we understand these linkages a little more, what does that mean?

49:14Like, why does this now open things up? Are we in a new environment? You kind of said, pre some of these unlocks and now are post. What does post look like? Yeah, this is a great question. So I think there's three big things that are sort of so -whats from this breakthrough. The first is that interpretability is now an engineering problem as opposed to an open -ended research problem. And that's a huge sort of sea change for the industry because up until now there were a number of hypotheses on how to interpret how these models were behaving and explain why, but it wasn't quite concrete, it wasn't quite understood which one of those approaches would work best to actually explain how these models work at very large scale, at frontier model scale.

49:56But I think this approach, this mechanistic interpretability approach and this paper that came out earlier this year shows that actually the relationships are so easily observable At a small scale that the bulk of the challenge now is to scale up this approach, which is an engineering challenge And I think that's massive because the engineering is largely a function of the resources and the investment that goes into scaling these models Whereas research can be fairly open ended and so I think one big takeaway from 2023 is that interpretability is gone from being a Research area to being an engineering area.

50:27I think the second is that if we actually can get this approach to work at scale then we can control these models. In the same way that if you understood how a kitchen made a dish and you wanted a different outcome, you wanted less lasagna and more pasta, and the next time you had them, the kitchen come together, now you can go to the Italian chef and say, could you please make that change next time around? And so that allows controllability. And that's really important because as these models get deployed and really important sort of mission critical situations like healthcare and finance and defense applications, You need to be able to control these models very precisely, which unfortunately today just isn't the case.

51:05We have very blunt tools to control these models, but nothing precise enough for those mission critical situations. So I think controllability is a big piece that this unlocks. And the third is sort of the byproduct of having controllability, which is once you can control these models, you can rely on them more. Increase reliability means not only good things for the customers and the users and developers using these models, but also from a policy and regulatory perspective, we can now have a very concrete, grounded debate about what models are safe and not, how to govern them, how to make sure that the space develops in a concrete and perically grounded way, as opposed to reasoning about these models in the abstract without a lot of evidence.

51:47I think one of the problems we've had as an industry is that because they hasn't been a concrete way to show or demonstrate that we understand these black boxes and how how they work, that a lot of the policy work and policy thinking is sort of worst case analysis. And worst case analysis can be fairly open to fear mongering and a ton of fud. And I think instead now we have an empirical basis to say, here are the real risks of these models. And here's how policy should address them. And I think that's a big improvement or a big advance as well for us all. Totally. I mean, it's huge. And it's kind of interesting because we don't know every little piece of physics, but we're able to deploy that in extremely effective ways and build all of the things around us through that understanding that has grown over time.

52:31And so it's really exciting that these early building blocks are getting in place. Maybe you can just speak to that engineering challenge or the flippinign that you said happened where we previously had a research challenge, which was somewhat TBD. When is this going to be unlocked? How is it going to be unlocked? And now we have, again, those early building blocks where we're now talking about scale. And I'll just read out a quick tweet from Chris, who I believe was on the Anthropic team. And he said, if you asked me a year ago, superposition would have been by far the reason I was most worried that mechanistic interpretability would hit a dead end.

53:05I'm now very optimistic. I go as far as saying it's now primarily an engineering problem, hard, but less fundamental risks. I think it captures what you were just mentioning, but maybe you can speak a little bit more to the papers and the scope that they've done this feature analysis within and what the steps would be to do this when we're talking about those much, much larger foundational models. So I think stepping back the way science in this space has done often is you start with a small, almost story like model of your problem. see if some solution is promising and then you decide to scale it up to a bigger and bigger level because if you can't get it to work at a really small scale, rarely do these systems work at large scale.

53:48And so I think, well, of course, the Holy Grail challenge with interpretability is explaining how frontier models that are the GPT -4 is in cloud 2s and barbs of the world, which are several hundred billion parameter in their scale. I think that one of the challenges we're trying to attack interpretability of those models directly is that they're so large and such complex systems, it is very untractable to try to tease apart all the different neurons in these models at that scale. And so I think what is the sort of classic scientific method of identifying a P3 dish at a a really small scale, getting a successful set of experiments to work at that scale, before then figuring out how to increase the scope of these experiments has largely worked well for AI.

54:37That's how scaling laws worked in the first place. We had GPT -2 before we had GPT -3 .5 because there was a much more tractable problem and scale to demonstrate that these models are capable of very high quality next token prediction. And I think that's what's with interpretability as well. And so the current breaks, who we have as an industry, have been demonstrated at this sort of toy prototype level with models that are in the tens of parameters. And I think the next step would be to figure out how to get these to work in the hundreds of millions of parameters, and then you could get to billions of parameters.

55:12And I think from the outside looking in, the state of interpretability can often seem underwhelming because these models are so small right now. But I think that's a little bit misleading because once we can get them to work at small scale, usually the industry has been fairly good at then getting to replicate those approaches at larger and larger scale. Now I should be clear, it's not easy. And there's tons of one -solve problems in the scaling part of this journey as well. Yeah, if I could just interrupt real quick, you mentioned the scaling laws and those have continued to scale, but we didn't necessarily know if that would be the case, it has, of course, proven to be the case as we move forward.

55:50But what are the challenges that you see that might be outstanding as we look to scale up some of this mechanistic interpretability research? What open challenges do you see on that path? To bar our analogy earlier of the kitchen, I think we as an industry now have a model of what's going on and some proof of what's going on with these features with a kitchen which has let's say three or four chefs. And so to figure out if this would work at frontier scale where you have thousands and thousands of chefs in each kitchen and in the case of a model you have you know billions of parameters, I think there are two big open problems that need to be solved in order for this current scale approach to work at scale.

56:31The first is increasing the autoencoder which is conceptually you can kind of think about as the model that makes sense of what's going on with each feature. And the autoencoder here is pretty small in the paper that came out in October. And so I think there's a big challenge where the researchers in the space have to figure out how to scale up the autoencoder in the order of magnitude of almost a hundred X expansion factor. And so that's a lot. And that's pretty difficult because training the underlying, the base model itself often requires hundreds and often billions of dollars worth of compute.

57:10And so I do think it's a fairly difficult and compute intensive challenge to scale the autoencoder. Now I think there's a ton of promising approaches and on how to do that scaling without needing tons and tons of compute, but those are pretty open -ended engineering problems. I think the second is to actually scale the interpretation. of these networks. And so as an example, if you find all the neurons and all the features related to, let's say pasta or Italian cuisine, and then you have a separate set of features that map to bass trees. Right. Now the question is, how do you answer a complex query and you ask the AI, hey, if I asked you a provocative question about whether people of a certain ethnicity, enjoy Italian cuisine or not, right?

57:57You need to figure out how those two features actually interact with each other at some meaningful scale. And that is a pretty difficult challenge to reason about too. And I think that's the second big open problem that the researchers call out in their work. And so I think the combinatorial complexity of each of those sets of features interacting with each other at the increasing scales is a non -linear increase in complexity that has to be interpreted. And so at least at the moment these are the two clear engineering problems that need to be solved, scaling up the auto encoder and scaling up the interpretation.

58:25But there probably are a list of long tail questions as well that I'm not addressing here. But those are sort of the two big ones that need to get all before this black box world has entirely moved to a transparent tier box world. Yeah, we've got a lot of steps along the way. And something that dawned on me as you were sharing that and potentially how compute intensive it might be, which also is very costly. Who is funding this? Like, obviously the paper we talked about is coming out of Anthropic. and other companies like OpenAI are also interested in understanding their technology better, but it's very easy to say, hey, we're putting together compute and economic resources to build a model because there's the consumer on the other side who uses that model.

59:06But what is the justification for folks to invest in building this mechanistic interpretability? Is it a policy thing or is it that actually having this interpretability makes the models better as well? Now, everybody who needs reliable AIs is essentially incredibly incentivized to invest in this area because as we were talking about earlier, if you are in the business of buying cakes from a kitchen and you don't like the cakes they're producing, but you can't tell them what to change about it and you can tell them how to improve the cakes, it can be pretty hard for you to get to the outcomes you want.

59:44And so I think what's happening right now in the industry is that interpretability research is largely done at labs that have the resources to do it and are incentivized to make their models more reliable and more steerable. I think there's a separate sort of body of work that's being done by academics and independent researchers, which is really important to support, which is safety research that is really hard for or folks outside of large labs to do, unless the open source ecosystem becomes more and more vibrant. I think these scaling problems are difficult to reason about unless you have access today to tons of compute to actually scale up the models and then study them.

1:00:29And I think one of the most exciting pieces of work that's being done around interpretability is by academic researchers, independent researchers who are reading what's happening at some of the large labs. and then trying to replicate that work on open source models. And so I think there are a number of really interesting experiments online by independent researchers where they've taken open source models like Lama 2 at 70b scale and I've started to try to get interpretability experiments to hold at that scale. And I think that's an increasingly important body of work because if we can't have the entire academic community and the independent researcher community participate in this interpretability research, we're going to end up in a situation where only a very limited number of people have access to doing that research.

1:01:14And so it's going to go slower almost definitionally. Yeah. So at the moment, the funding and the resources are coming from the largest labs. And I think there's a glimmer of hope where more and more open source work is being done, but we've got to make sure that continues and grows in volume and doesn't get paused. Completely. Yeah. And maybe one other way to ask the question that I just asked is what is this all unlock, right? If we have these new tools, if we turn Black Box to Clear Box, how does this change the game? And maybe you could speak to what you're excited for specifically coming into 2024 as it relates to mechanistic interpretability.

1:01:46Yeah, so to be clear, I'm excited about all kinds of interpretability or at explainability. I'm broadly very excited about 2024 as the first time or at least the year where the most amount of interest and attention is being paid to explainability. The last few years, the attention was all on the how and the what. People are just incredulous at the capabilities of these models. Can we get them to be smarter? Can we get them to reason about entirely new topics that maybe weren't in the original pre -training data set? And that's been totally reasonable. But I think more attention on the why of these models and to explain how they work.

1:02:23It has been the big blocker on these models getting deployed outside of just a few consumer use cases where the costs of the model not being as reliable, as steerable are low. And so low precision environments, consumer use cases, where people are more forgiving and tolerant of mistakes by the model and so on, is largely where the bulk of the value has been generated in AI today. But I think if you want to see these models take over some of the most impactful parts of our lives that they currently aren't deployed in, things like healthcare, things like financial parts of the world where a lot of really tedious work is being done by folks who would love to have these models automate a lot of our decisions, I think that those mission -critical situations require a lot more reliability and predictability.

1:03:12And that's what interpretability ultimately unlocks. If you can explain why the kitchen does something, then you can control what it does. And that makes it much more reliable. And therefore, it's going to be used in more and more situations and in more use cases and then more and more and more impactful customer journeys where today a lot of the models don't actually make the cut. No, it's so true. Actually, something that also just dawned on me as you were talking is almost everything in this world has margin for error. Right. There is error inherently. And in most things, however, if you can understand, if you can explain that error and constrain it to something that other people can get behind, it's just much more likely that people will want to engage with that thing because they can at least understand what is coming out of it.

1:03:56And so, yeah, I feel like that picture is very compelling and I hope we can get there. I hope so too. I think to be clear, we're not there yet, but we've got the glimmers now of approaches that might work. And I think what I'm excited about 2024 is a lot more investment, a lot more energy, a lot more of the best researchers in this space spending their time on interpretability. Yeah. Well, we have some of the smartest people in the world working on AI and we saw how quickly things moved in 2022, 2023, so hopefully in 2024 some of this interpretability work moves just as quickly. I hope so. I'll wear my fingers crossed.

1:04:29All right, I hope you enjoyed these three big ideas. We've got a lot more on the way, including the new age of maritime exploration that takes advantage of AI and commuter vision, AIR for schemes that never end, and whether VoiceFirst Apps may finally be having their moment. If you can't wait and want to see our full list of 40 plus big ideas today, you can head on over to a16z .com slash big ideas 2024. It's time to build.