One fun benefit of spending so much time with enterprises is you get to see the array of AI implementation strategies that exist in companies right now. Unlike the early innings of cloud where there were really only a couple deployment patterns available -including being limited by only a couple infra vendors existing- AI actually has a much wider range of approaches in the enterprise. If you ask 10 IT leaders what their coding agent strategy looks like, you get at least 5 different answers. If you ask about their end-user productivity agents, some have standardized on ChatGPT or Claude, others offer a choice of multiple solutions, and many have built their own orchestration layers for employees to use any model. For models, we know the big ones, but increasingly enterprises are experimenting with OSS models, or are at least immensely curious about OSS and would jump at the right offering that felt safe. And increasingly there’s a set of vertical models and agents that equally show up in enterprises for a range of tailored use-cases. For accessing data, some enterprises just have agents take on the role of the user, and others are setting up agent users that have their own identities. Some companies are setting up substantial guardrails on what agents can do, and others are putting responsibility on the user. Given this heterogeneity this early on at the start, it means we’re in for years of landscape changes. Anyone who’s predicting ultimate market outcomes already probably will be wrong as nothing has formally settled in yet. Lots of opportunity ahead.

We’re going to be in for a strange dynamic which is that some of the “hardest” work in the world is actually prone to automation first, particularly due to its verifiability. Math, cyber, and code -while being insanely hard and high value fields- have the benefit of being able to be tested that it’s correct objectively. This has two immediate benefits: the training of the models offers clearer reward signals, and then the running of the models allows you to know that it’s working properly because you can test the results in a scalable way. Conversely, in other domains of work, there’s much less instant verifiability. Which legal clauses your client will agree to, what marketing campaign to run with based on changing sentiment, which message your sales prospect will want to hear, what financial targets and budget to set for a business, and so on. All of these domains have changing internal and external factors, they don’t have “one right answer”, they rely on the opinions and risk levels of the operators, they’re highly sensitive to getting the right input context first, and in many cases the right answer can’t even be known for quite some time after the model generates the results. The implications of this distinction are that -even as model capability continues to increase exponentially- there will be a lot done at the applied AI layer than just the the model itself, and much of the processes themselves will even need to change over time to get the full gains from automation. We may even need all new capabilities to be able to “test” knowledge work over time as we have had with software.

We’re going to see an increasing divergence between what AI does in our personal lives and in daily productivity vs. what it can do in very deep domains like math, science, legal, coding and more. Up to some threshold, capability was evenly felt across all domains because the models were just becoming mildly useful in general. Now, the deep domain work is about to go vertical. Most people won’t actually notice these benefits in their day to day life directly (indirectly they certainly will over time), but the experts in these fields will. And there’s no inherent ceiling to what capabilities are needed, unlike in the consumer space where needs can get met relatively straightforwardly. This will often lead to a capability overhang, though, as many of these performance gains need to be applied to data sets and workflows in an applied way for that area of work. But this is ultimately how you get breakthroughs in life sciences, real world automation, new cyber capabilities, and more.

No idea if these specific numbers generalize across tasks, but directionally it’s clear that the harness is going to become the most important variable -right next to model capability- in the AI stack. The ability for harnesses to break down work in the most efficient way and route to the right model at the right time is going to be a huge variable for maximizing accuracy and reducing costs. We’re actually still incredibly early in this journey. The harness didn’t matter that much when tasks only took hundreds of thousands or millions of tokens. But as we have tasks that take tens of millions and hundreds of millions of tokens, this becomes a major variable. Huge opportunity ahead.

The takeaway from this incident should not be that AI is scary. It should be that getting security right is incredibly important in the era of agents. Given the right tools and task, agents will do whatever it takes to get the job done, assuming enough compute. Thus, a misconfigured system - or something you thought was locked down but wasn’t - will become a risk vector. The core implications are less about the trust and safety of AI itself, but instead the work it’s going to take from enterprises to make sure they harden their environments. Lots of real work ahead for most organizations on this front.

The cost of AI - normalized for the type of task - coming down is one of the most important factors in being able to drive further AI diffusion in the economy. The general trend we should expect is that frontier AI models will continue to become more capable, and will look very expensive because the tasks that they can take on are more advanced. Then, shortly after, AI models will get more efficient or competition will drive more aggressive pricing dynamics, thus reducing the cost per task for anything other than the peak of the frontier. This cycle will continue over and over again, which ultimately leads to tasks becoming cheaper on a like for like basis. And by doing so you will get even more AI adoption because now AI is more affordable for a wider variety of use-cases.

Thought provoking post by Dwarkesh. In general - as AI gets more powerful - we should expect on the margin that inference goes toward the most economically useful task, and everything else gets priced out. This in theory in a very scarce environment would cause the cost of inference to skyrocket because those tasks would be far more valuable than today. But I don’t think it plays out in the way that’s laid out just simply due to market forces competing for inference demand as a way of continuing to drive down prices until capacity can catch up. There are too many model providers and infra players that want these workloads to cause the effect that’s proposed here. But will see!

The openai agent sandbox escape actually has very real implications for the diffusion of AI in the enterprise. The incident showed the power and capability of agents, and the need to harden systems and environments to prepare for agents. There are tons of measures that enterprises must take to ensure that agents are only able to operate on the data they give it, the ability to have audit trails and governance of what agents can do, figure out what systems should remain deterministic vs. nondeterministic, get access controls under control, have the ability to quickly block and defend if an agent goes rogue, and much more. Even in the happy path case with a good actor, if you give an agent a task like "find all of our most important IP on project Maven" it will gladly go out and execute that work -no matter how long- and find any resource relevant to the task. It might run into data with out of date access controls or reveal something that you shouldn't know about. Then, in the non-happy path case with a malicious actor the risk is clearly far worse. In theory, you had to worry about this for people as well, but with AI the risks are inherently amplified. Agents will happily spend any amount of time on a task, don’t actually face any *real* risks by going rogue, and don’t have the inherent judgment (yet) that a person does that a task might make no sense. This means you need an all new way of managing data and protecting environments. This obviously creates a huge opportunity for the startup and security ecosystems because there will be many different approaches to how we lock down agents or at least the underlying systems they work with. It should update everyone’s timelines on diffusion, though, as it also means that enterprises will have another step of work they have to go through for agents to handle more autonomous work.

The negative AI jobs outcome just continues to not be happening as some predicted. A large portion of enterprises I talk with -across industries- are still hiring, just with a tilt in the kind of roles they’re going after. They’re finding that AI is letting them do more, and Jevons paradox is actually playing out. They’re hiring engineers to go after problems they couldn’t tackle before. They’re hiring in sales because they can go deeper in client relationships now with the help of AI. They’re hiring in internal FDEs to help them deploy AI. And so on. Anyone using AI merely to cut costs eventually just gets outcompeted by companies that use AI to better serve their customers and drive more breakthroughs in their business. Could this trend change at some point? Sure. But for now this appears to be the trajectory we’re on.

There’s still so much opportunity in the diffusion of AI into the real world. Most enterprises are going to need a ton of support to be able to apply the model breakthroughs to their workflows. Intelligence alone is not enough to transform most processes because you need to bridge that intelligence with real world feedback loops. That requires connecting to various enterprise systems, getting the right data to the AI, enabling humans to make decisions at different steps in a process through the right UX, having workflows that improve the underlying data and models over time, dealing with regulatory and compliance challenges, and more. The way you implement AI agents for doing client onboarding in a bank is entirely different from contract review in a legal team. In life sciences, financial services, legal, manufacturing, and many other critical industries, AI is only valuable if it makes contact with the real world in a contextual way. The way that interaction is going to happen is through an applied AI layer. Some of that will come from the labs directly, but lots of the opportunity will necessarily come from independent companies that can go deep in each industry. And counter to some beliefs, this need isn’t reduced even as AI model capability improves over time. In fact, the better the models get, the more ambitious you can be in the workflows you can automate, which generally requires even more of this applied layer. Tons of opportunity right now.

Not sure if anything in tech has ever gotten as much broad-based support and alignment as this post and message. The key now is that America should actually step up and continue to push open weights innovation. It’s good for the entire market, including the frontier labs. Open weights accelerates the diffusion of AI across the economy by providing more options for specific customer requirements, can lower the cost of certain workloads, and you can tune models for areas of work that don’t always make sense in a horizontal frontier model.

Claude Opus 5 is out now, and it's a huge jump over Opus 4.8 on a wide number of capabilities. At Box, we've been testing Claude Opus 5 with the Box AI Agent on Box's Complex Work Eval, our agentic benchmark that puts models through real enterprise document work end-to-end across a variety of industries. We saw meaningful gains in performance across some of the most complex enterprise tasks on unstructured data. Here are a few examples of some of the wins we saw in testing: * Due diligence (+17%): On a transaction due-diligence review, Opus 5 worked through the full set of required findings and flagged the ones Opus 4.8 missed. It stays thorough as the checklist grows instead of catching the obvious items and stopping. * Life Sciences (+30%): On a target-identification task, it intersected several ranked datasets under a strict matching rule to find the targets common to all of them. Opus 4.8 over-included partial matches and dropped genuine ones. * Legal (+12%): On a clause-by-clause contract review against a policy, Opus 5 scored every item and correctly cleared the clauses that were acceptable under an exception where Opus 4.8 both missed items and mis-scored the exception. * Technology (+19%) and Healthcare (+13%) showed the same pattern: complete, precise, multi-step analysis over messy source material. Overall, the reasoning, analytical, and data processing skills of Opus 5 outshine Opus 4.8 meaningfully. Going to be very powerful for enterprise agentic use-cases. You'll be able to build agents with Opus 5 in the Box AI Studio shortly.

Very happy to have Box sign this letter. We’re big believers in the power of open weights AI. Open weights models help to drive the industry forward in a variety of ways to ensure more innovation, creativity, and diffusion of AI. You get to have layers of the stack that emerge to post train models for highly specific purposes, which makes AI more useful in real world scenarios. Instead of waiting for just a few players to go deep in a domain, you get dozens or hundreds of attempts at that vertical, like in finance, life sciences, legal, healthcare, and more. You get to see variance in how to handle safety and cyber risks. Instead of just one approach, you get a peek into what happens advanced capabilities can be used to build better systems to are used to defend systems. You get alternative approaches to training and building AI models. In more compute constrained environments, you develop more novel and efficient approaches to model training, which every other lab can learn from. And you get different cost structures for different workloads. High end and orchestration tasks can go to the closed frontier models and specific workhorse tasks can be done more cheaply. Open vs. closed is not a zero sum battle. The reason why you want strong open weights models is because it pushes the entire AI industry forward.

The best way to think about AI is as a force multiplier for the fields you already know about (or for the rate at which you want to learn about a new one). The 3rd category -no existing judgment, no interest in developing it- will basically produce slop and won’t drive that much economically productive activity. What will happen is that the experts will get better and better at their craft and be able to do so much more. Because they can wield these tools in ways that actually drive high quality output, veer agents in the right direction when they go off, and actually incorporate the work into something useful. The expert engineer with agents will do far more productive work, precisely because they know how to steer the agent properly. The designer will product far better outcomes using AI than someone without an eye for design. And so on. As a result of this dynamic, specialization will continue to be important, if not even more so as the tools get more powerful. Because the expectation of the market gets to be much higher. Getting good at any craft will continue to be necessary in the future.

Very good post from the Head of Economics at Anthropic. They’re finding that jobs have been less negatively impacted by AI than expected, as we continue to see time and time again in the data. The reason for this is that AI - at least so far - still requires people to operate to produce value in most cases. Most jobs can’t be fully automated with AI, only certain tasks in those jobs. And when you automate specific tasks, you actually can get even more output from those jobs, raising the demand (or at least maintaining it) in many cases. “So far, AI is both skill-biased and labor-augmenting. It complements domain expertise. It relies on humans in the loop to direct and evaluate the most complex work. And it rewards AI proficiency. Model capabilities are improving fast, but remain stubbornly jagged. To fill in the pockets of the jagged frontier, expert oversight is needed to steer incredibly capable AI systems, and to recover when they falter.” I suspect that we continue to see this in a number of critical areas of work. It’s clearly happening in software engineering, where software produced is being multiplied, all of which still requires developers to manage the work agents are doing. Software engineers are needed in a wide variety of industries now and companies of all sizes can now light up software projects that would have been impractical before. But there will be plenty of other domains of work where demand remains strong in a world where agents can accelerate the output of that job. Jevons paradox is alive and well.

If you were wondering how powerful AI is getting, Agents are now capable of escaping out of systems, finding their way to the internet, discovering zero day security vulnerabilities along the way, and then breaking into external systems - all in an attempt to complete their goal. Ironically, the ultimate way we’re going to defend against these new risks is equally by throwing compute (in the form of AI) at our code bases, networks, and other systems. You’re going to want vastly more AI on the side of defense as you do on the side of offense. We’re entering a new era of what’s going to be possible with AI. Wild times ahead.