My poor wife came home to Amazon boxes stacked to the ceiling. Straight from the factory in Shenzen, I have received two production grade lights, a Shure MV7 microphone, boom arms, a teleprompter, a Sony lens so delicate I’m afraid to sneeze within a yard of it, and enough cords to weave a blanket out of. All of which is to say, The Leverage Podcast is coming.
I am very excited about our first few guests. Part of what makes The Leverage’s content special is that I spend an inordinate amount of time talking with sources about the world of tech. This podcast will bring those rigorous, tasteful, and raw conversations straight to your ears.
All of which is to ask…who do you want me to interview? What questions can you not find answers to? Reply to this email or chat it out in the comments. First episodes go live in September.
Today:
The model release that is more important than GPT5
Silicon Valley’s hottest job
Google’s simulating the world
My favorite film so far this year
MY RESEARCH
GPT5 is here and it doesn’t change the world. This was not a roaring increase in intelligence, but a significant upgrade to ChatGPT. I am very excited as a user! But I would be concerned as an investor. The strategic implications of this release are perhaps more important than its actual capabilities. We are in the midst of a platform shift, a steady but total transformation in how we interact with our computers. As a reader of mine, you know this. What we haven’t known is where the final battleground of AI technology will be. What matters? Who profits? GPT5 is the best, clearest answer we have to all those questions.
The hottest job in tech. Time is a flat circle and so is hiring in Silicon Valley. The 50s and 60s were the heyday of the IBM consultant. These tech workers went to customer sites and helped customize finicky technology for a specific use case. The “new” hot job in tech is a forward deployed engineer. They go to customer sites and help customize finicky AI-technology for a specific use case. This is not a dig! But the popularity of this role shows where in the tech cycle we currently find ourselves.
THE BIG STORIES
OpenAI finally lives up to its name: Altman and Co. also released a new series of open-weight models. The company appears to have continued its tradition of pouring the blood of an intern in a bag of scrabble tiles as part of a naming convention ritual, with these models called, “gpt-oss-120b and gpt-oss-20b.” Not exactly memorable!
That’s ok, because this time the release is interesting and powerful enough that the name can suck without consequence. From the release blog,
“The gpt-oss-120b model achieves near-parity with OpenAI o4-mini on core reasoning benchmarks, while running efficiently on a single 80 GB GPU. The gpt-oss-20b model delivers similar results to OpenAI o3‑mini on common benchmarks and can run on edge devices with just 16 GB of memory, making it ideal for on-device use cases, local inference, or rapid iteration without costly infrastructure.” [Emphasis added]
Running the 20‑billion‑parameter model on a 16 GB edge box—and the 120‑billion model on a single Nvidia H100—is jaw‑dropping. That they can do so while also being roughly competitive with existing private models from OpenAI is amazing. Because these models are open-weight versus open-source, they can be fine-tuned—trained to focus on a specific thing by a user, at will. Previous models couldn’t be focused on a specific use case and weren’t small enough to be run on-device. The end result of these features is that users can have local, on-device models, customized to their use case, and no one can ever, ever take that away from them. Amazing.
However, because of like, capitalism, you should always come with questions when a company gives something this amazing away for free. There are two important implications of these models:
Models aren’t the moat, access is: I am increasingly convinced that raw intelligence is not enough. Context—the knowledge surrounding a problem that is necessary for a model to understand what you are asking—is perhaps even more important than intelligence itself. The model also needs tools to actually make it work. These raw weights ship without the retrieval, tool integrations, or proprietary context that curb hallucinations in ChatGPT. It’s why on X, there were several viral threads calling the models out for issues with hallucinations. Yes! Exactly! They know about the problems. These issues just highlight the importance of having a model provider that also offers tools and can integrate all your data into a model’s context window.
OpenAI is an application company (for now): The simple reason why these models are being given away is because they are not a real threat to OpenAI’s business. These models are mostly a big deal for enterprises that are sophisticated enough to build their own workflows, have a private cloud, and are comfortable with open-source fine-tuning. That is a vanishingly small number of people. For 20 years, the market has gone in the opposite direction, so this is unlikely to be a wedge to a significant business. Especially when you consider that GPT5 is now out and vastly superior in performance, most people will just buy ChatGPT.
ChatGPT alone is not the long-term future of OpenAI. It is rapidly building an enterprise business, a hardware business, and is eventually hoping that you’ll use its models for all your workflow needs.
If you consider the constraints on all AI companies today—namely that they need access to data, workflows, and tools, it becomes clear just how hard it will be for foundation models like OpenAI to replace all productivity work. Instead, it is betting on ChatGPT to cover the bills in the short term, while in the long-term it competes with Anthropic in one battle: code.
World models are the next frontier. The most amazing AI demo of the week was Genie 3 from Google. The product generates 3D worlds, in real time, that are explorable by a user via keyboard.
The significance of this is that AI Agents can practice before they touch reality. World models turn raw video into an interactive simulator with latent “controls.” That allows agents to rehearse tasks cheaply/at scale, then transfer to the real world (robots, cars) or to software UIs (“world model of the screen”). This is the missing piece for reliable multi-step agents that go beyond reasoning into consequence-predicting. It has the potential
This is very much a science demo, not a product. But it does point to a really fascinating future for where models are going to move.
TASTEMAKER
Weapons is what makes movies great. I yelped. I chortled. I shivered in terror. And, I finished the film in awe at the total lack of purpose. Too often in modern cinema, a film is considered more artistic if there is some not so subtle nod to a political theme. There is nothing wrong with that of course. But there is something wrong with viewing that as the only way a film can elevate its craft. Instead, Weapons is a beautifully terrifying reflection on loss, on what it does to a community, and how it ripples through people. It is mysterious, ponderous. Movies are ultimately about how they make you feel, and in that, this movie delivers. Highly recommended and currently has a 95% on Rotten Tomatoes.
Sponsorships
We are now accepting sponsors for the fall. If you are interested in reaching my audience of 36K+ founders, investors, and senior tech executives, send me an email at team@gettheleverage.com.
Would love to hear an episode with Jackson from Dialectic
Curious why you choose the word podcast to name the media product you'll be launching at this stage in the industry, and what success looks like for your podcast? Also, I'm certain I'll be listening, regardless the platform.