For my last issue of the year, I\u2019m focusing on the AI talent war, which is a theme I\u2019ve been covering since this newsletter launched almost two years ago. And keep reading for the latest from inside Google and Meta this week.<\/p>\n<\/div>\n
But first, I need your questions for a mailbag issue I\u2019m planning for my first issue of 2025. You can submit questions via this form<\/a> or leave them in the comments.<\/p>\n<\/div>\n This week, Databricks announced<\/a> the largest known funding round for any private tech company in history. The AI enterprise firm is in the final stretch of raising $10 billion, almost all of which is going to go to buying back vested employee stock.<\/p>\n<\/div>\n How companies approach compensation is often undercovered in the tech industry, even though the strategies play a crucial role in determining which company gets ahead faster. Nowhere is this dynamic as intense as the war for AI talent, as I\u2019ve covered before.\u00a0<\/p>\n<\/div>\n To better understand what\u2019s driving the state of play going into 2025, this week I spoke with Naveen Rao<\/strong>, VP of AI at Databricks. Rao is one of my favorite people to talk to about the AI industry. He\u2019s deeply technical but also business-minded, having successfully sold multiple startups. His last company, MosaicML, sold to Databricks for $1.3 billion in 2023. Now, he oversees the AI products for Databricks and is closely involved with its recruiting efforts for top talent.<\/p>\n<\/div>\n Our conversation below touches on the logic behind Databricks\u2019s massive funding round, what specific AI talent remains scarce, why he thinks AGI is not imminent, and more.<\/p>\n<\/div>\n The following conversation has been edited for length and clarity:<\/em><\/p>\n<\/div>\n Why is this round mostly to help employees sell stock? Because $10 billion is a lot. You can do a lot with that.<\/strong><\/p>\n<\/div>\n The company is a little over 11 years old. There have been employees that have been here for a long time. This is a way to get them liquidity.\u00a0<\/p>\n<\/div>\n Most people don\u2019t understand that this is not going into the balance sheet of Databricks. This is largely going to provide liquidity for past employees, [and] liquidity going forward for current and new employees. It ends up being neutral on dilution because they\u2019re shares that already exist. They\u2019ve been allocated to employees and this allows them to sell those to cover the tax associated with those shares.<\/p>\n<\/div>\n How much of the rapid increases in AI company valuations have to do with the talent war?<\/strong><\/p>\n<\/div>\n It\u2019s real. The key thing here is that it\u2019s not just pure AI talent \u2014 people who come up with the next big thing, the next big paper. We are definitely trying to hire those people. There is an entire infrastructure of software and cloud that needs to be built to support those things. When you build a model and you want to scale it, that actually is not AI talent, per se. It\u2019s infrastructure talent.\u00a0<\/p>\n<\/div>\n The perceived bubble that we\u2019re in around AI has created an environment where all of those talents are getting recruited heavily. We need to stay competitive.\u00a0<\/p>\n<\/div>\n Who is being the most aggressive with setting market rates for AI talent?<\/strong><\/p>\n<\/div>\n OpenAI is certainly there. Anthropic. Amazon. Google. Meta. xAI. Microsoft. We\u2019re in constant competition with all of these companies.<\/p>\n<\/div>\n Would you put the number of researchers who can build a new frontier model under 1,000?<\/strong><\/p>\n<\/div>\n Yeah. That\u2019s why the talent war is so hot. The leverage that a researcher has in an organization is unprecedented. One researcher\u2019s ideas can completely change the product. That\u2019s kind of new. In semiconductors, people who came up with a new transistor architecture had that kind of leverage.\u00a0<\/p>\n<\/div>\n That\u2019s why these researchers are so sought after. Somebody who comes up with the next big idea and the next big unlock can have a massive influence on the ability of a company to win.<\/p>\n<\/div>\n Do you see that talent pool expanding in the near future or is it going to stay constrained?\u00a0<\/strong><\/p>\n<\/div>\n I see some aspects of the pool expanding. Being able to build the appropriate infrastructure and manage it, those roles are expanding. The top-tier researcher side is the hard part. It\u2019s like looking for LeBron James<\/strong>. There are just not very many humans who are capable of that.\u00a0<\/p>\n<\/div>\n I would say the Inflection-style acquisitions<\/a> were largely driven by this kind of mentality. You have these concentrations of top-tier talent in these startups and it sounds ridiculous how much people pay. But it\u2019s not ridiculous. I think that\u2019s why you see Google hiring back Noam Shazeer<\/strong><\/a>. It\u2019s very hard to find another Noam Shazeer<\/strong>.\u00a0<\/p>\n<\/div>\n A guy we had at my previous company that I started, Nervana, is arguably the best GPU programmer in the world. He\u2019s at OpenAI now. Every inference that happens on an OpenAI model is running through his code. You start computing the downstream cost and it\u2019s like, \u201cHoly shit, this one guy saved us $4 billion.\u201d<\/p>\n<\/div>\n \u201cYou start computing the downstream cost and it\u2019s like, \u2018Holy shit, this one guy saved us $4 billion.\u2019\u201d<\/p>\n<\/div>\n<\/div>\n What\u2019s the edge you have when you\u2019re trying to hire a researcher to Databricks?<\/strong><\/p>\n<\/div>\n You start to see some selection bias of different candidates. Some are AGI or bust, and that\u2019s okay. It\u2019s a great motivation for some of the smartest people out there. We think we\u2019re going to get to AGI through building products. When people use technology, it gets better. That\u2019s part of our pitch.\u00a0<\/p>\n<\/div>\n AI is in a massive growth base but it\u2019s also hit peak hype and is on the way down the Gartner hype curve<\/a>. I think we\u2019re on that downward slope right now, whereas Databricks has established a very strong business. That\u2019s very attractive to some because I don\u2019t think we\u2019re so susceptible to the hype.<\/p>\n<\/div>\n Do the researchers you talk to really believe that AGI is right around the corner? Is there any consensus of when it\u2019s coming?\u00a0<\/strong><\/p>\n<\/div>\n Honestly, there\u2019s not a great consensus. I\u2019ve been in this field for a very long time and I\u2019ve been pretty vocal in saying that it\u2019s not right around the corner. The large language model is a great piece of technology. It has massive amounts of economic uplift and efficiencies that can be gained by building great products around it. But it\u2019s not the spirit of what we used to call AGI, which was human or even animal-like intelligence.<\/p>\n<\/div>\n These things are not creating magical intelligence. They\u2019re able to slice up the space that we\u2019re calling facts and patterns more easily. It\u2019s not the same as building a causal learner. They don\u2019t really understand how the world works.\u00a0<\/p>\n<\/div>\n You may have seen Ilya Sutskever\u2019s<\/strong> talk<\/a>. We\u2019re all kind of groping in the dark. Scaling was a big unlock. It was natural for a lot of people to feel enthusiastic about that. It turns out that we weren\u2019t solving the right problem.<\/p>\n<\/div>\n Is the new idea that\u2019s going to get to AGI the test-time compute or \u201creasoning\u201d approach?<\/strong><\/p>\n<\/div>\n No. I think it\u2019s going to be an important thing for performance. We can improve the quality of answers, probably reduce the probability of hallucinations, and increase the probability of having responses that are grounded in fact. It\u2019s definitely a positive for the field. But is it going to solve the fundamental problem of the spirit of AGI? I don\u2019t believe so. I\u2019m happy to be wrong, too.<\/p>\n<\/div>\n Do you agree with the sentiment that there\u2019s a lot of room to build more good products with existing models, since they are so capable but still constrained by compute and access?<\/strong><\/p>\n<\/div>\n Yeah. Meta started years later than OpenAI and Anthropic and they basically caught up, and xAI caught up extremely fast. I think it\u2019s because the rate of improvement has essentially stopped.<\/p>\n<\/div>\n Nilay Patel compares the AI model race to early Bluetooth. Everyone keeps saying there\u2019s a fancier Bluetooth but my phone still won\u2019t connect.<\/strong><\/p>\n<\/div>\n You see this with every product cycle. The first few versions of the iPhone were drastically better than the previous versions. Now, I can\u2019t tell the difference between a three-year-old phone and a new phone.\u00a0<\/p>\n<\/div>\n I think that\u2019s what we see here. How we utilize these LLMs and the distribution that has been built into them to solve business problems is the next frontier.\u00a0<\/p>\n<\/div>\n A few notable moves this week:<\/em><\/p>\n<\/div>\n If you aren\u2019t already getting new issues of Command Line<\/em>, don\u2019t forget to subscribe to The Verge<\/em><\/a>, which includes unlimited access to all of our stories and an improved ad experience on the web. You\u2019ll also get access to the full archive<\/a> of past issues.<\/p>\n<\/div>\n As always, I want to hear from you, especially if you have a tip or feedback. Respond here, and I\u2019ll get back to you, or ping me securely on Signal<\/a>.<\/p>\n<\/div>\n<\/div>\n\u201cIt\u2019s like looking for LeBron James\u201d<\/h3>\n<\/p>\n
Elsewhere<\/h3>\n<\/p>\n
\n
What else you should know about<\/h3>\n<\/p>\n
\n
Job board<\/h3>\n<\/p>\n
\n
More links<\/h3>\n<\/p>\n
\n