OpenAI podcast mentions

Quote

at the right time. The problem ha w what happens is a at scale across the enterprise when you have like hundreds and thousands of developers, not everyone is using the tool. uh with the same level of efficacy. So all of the CI tools, um codec, plot code. you name it, require a significant amount of uh setup. that work for anthropic don't work for open AI. Like for example, open AI, um Don't use XML tokens in their training, but Anthropic does, right? So if you use XML tokens, with enth with open air or if you shout in your prompts, which you have to do with cloud at times. You have to shout at Claude to get it clean. It's widely held that uh GPT five point three gets many things. That opus does not, right? So there's all of these complex agentic engineering So that's the part that is agentic engineering where you recruit the right agent with the right set of prompts and tools with the right level of prompt engineering for the right task, right? Because there are definite tasks that GPT is better than Opus for. And then there's context engineering. which optimizes for giving the agent the right amount of information at the right time and f getting it to focus on the the smallest possible task that is efficient for that.

Quote

You know, there's no restrictions on what you can use. I believe VS Code is similar. You know, you can use whatever models. you want. I mean personally, just to finish the long thread, I'm excited about local models. I think Those are really, really interesting. So you can already certainly in PyCharm and Jeprains ones, you can can hook in AI chat and features into local models or you could just use Google Deep Seek has an eight billion parameter model. OpenAI, I'm blanking on the name. But then you can even do things like there's a six hundred seventy one billion parameter. deep seek model that you can run with OLAMA cloud. So this is the idea that they'll run the inference. for you and charge you presumably less than if it was a proprietary model. That's super, super interesting. The idea that you can have almost frontier level, presumably secure, less cost. Yeah long long answer to a short question but JeffBrains is adapting with wherever whatever helps developers. And I suspect there'll be some role for an IDE, but what we call an IDE will probably look A lot different in five years. I think too that one of the patterns that are starting to develop around the use of these agentic development managing the parallel work streams where one of those aspects is the

Quote

Um Intel was baked into this GPT or baked into this project and we couldn't really surface it or we couldn't really like adapt it in any So in some ways we were actually flying more blind using ChatGPT as found. Yeah, and you know we were actually talking about part of the the call about um chat GPT stuff and maybe we'll get into that uh later in regards to the uh the inevitable downfall of open AI. No, just kidding. Um but how when you you're going about building a project I know you know we're really excited to see what you're doing as this is about to launch. you go about figuring out what is a priority to people, right? there already. What was kind of your approach on figuring out how to fill out Yeah. have a bunch of um I guess founder friends or networks around us that that we leveraged. So we started doing a lot of um customer interviews, customer discovery. validation type uh work and we got a a range of different things. I think one of the the patterns we saw is most founders don't even know what they need help with. And they have like sometimes very superficial things like, oh I need help with fun

Quote

um uh adapters is what we call them and for example it Each uh provider has a lot of adapters depending on what they support. So for example For example, um open AI has a video adapter, image adapter, text adapter. Yada yada. And then you just give it to the appropriate method from tenstake. So for example if we want to chat, you say, Okay, let me grab the chat method, let me grab the open AI adapter, put them together, I'm talking with OpenAI model. And it it works like that. So you want to generate an image. I grab the image adapter, I grab um the generate image method, I put them together. I'm generating images. So you can build anything like from simple Like simple scenarios to something like T Tree chat, it's really up to you what you So there's a lot of folks who are not going to be able to do that. operating in this space who are trying to build these like bridges and I think they're they're going out After the same sorts of problems is like you have all the providers and their APIs are different and you kinda want a cohesive interface. where you can like switch providers like relatively easily and not get locked. I I know that like Vercel and the U.S. particular has like a very popular um SDK um and like

Quote

mindset to things and constantly be asking where is the Asian making. mistakes? Where am I spending my time? How can I not spend that time going forward forward and then build confidence in the automation that I'm putting in place so I have salt. All right, we're in the studio with Ryan Lopoplo. from OpenAI. Welcome. Hi. Thanks for visiting San Francisco and thanks for spending some time with us. Yeah. Thank you. I'm super excited to be here. You wrote a blogbuster article on harness engineering. It's probably Thank you. Is it fun to feel like we've defined the discourse in some sense? Let's contextualize it. bit on this first podcast you've ever done. Yes. And thank you for spending it with us. W what is where is this coming from? What team are you in? All that jazz. Sure, sure. I work on frontier product exploration. new product development in the space of OpenAI Frontier, which is our enterprise platform. for deploying agents safely at scale with good governance in Any business. And the role of me and my team has been to figure out And you have a background, I'll just squeeze it in there.

Quote

Has had a major security incident because of an AI tool that one of their employees employees was using and the FTC just dropped a settlement with Clarify over the company using three million okay Cupid photos to train facial recognition. Snapchat laid off around a thousand people and said that it is part of their AI initiative because AI is writing more than 65% of their new code. They dropped a main major new image model over at OpenAI called image two point oh that is a huge Step up. I've been playing with it today. We'll talk about that. And in addition, anthropics myth of. Ghost model has of course guinea is starting to get a bad rep and Sam Altman is going and calling them out. for their fear based marketing. They also signed a huge enterprise deal at Novo Nordisk, the Ozempic company we'll be talking about. And I think one of the biggest stories is that M Amazon is pouring over twenty-five billion dollars into Anthropic on top. Now if you're still paying for Chat GPT, Claude, Gemini, Grok, any of the audio models, 11 labs for audio. and any of the image models, I've got to tell you about AI Box. This is what I personally have built It's what I'm recommending to my friends who ask me how to actually use AI without going broke on the system. subscriptions, you get access to over 80 different AI models in one place. All of the top models

Quote

Oh. For for sure, for sure. Um a few other things that you know we haven't mentioned on this podcast yet is uh openai came out with their GPT five point three uh codex sp Spark model. So I I know they were trying to uh simplify their naming with G. But it seems like whoever's in charge of uh uh naming things at open AI. AI uh is seems to be still working there. So yes, we have the G And normally I wouldn't talk about this kind of model, but I think it's really interesting because it is their first deal with Cerebris. So cerebris, for those that don't know, is a company that makes inference chips. So they make really, really big uh physical computer chips like The biggest chip that you can possibly put um that you can possibly make. uh and it's all just like ai specific. So these chips are not something you could go put in like your laptop. Um like the si the chip is like the size of a like a dinner plate. And um it's specific for AI. It costs like a million plus.

Quote

But the overall project is is much more important than it than that individual uh dependency and that individual dependency could be swapped out for any number of things in in reality. I think I think That is one of the things that we're seeing now. You know, we've talked about twenty twenty six being, you know, with Claude Code and and others of codecs from OpenAI and then uh as well as even open source models that are that are available for coding and There's some really good ones still. And so that's but one of the things that I or make your systems more dynamic because you can you can take major components and um and redo interfaces and stuff like that. I've seen value added there. But it really calls out the fact that the while that has changed in that while some of the tools that we have now are speeding things up, it still takes getting maybe a novel idea or maybe just an iteratively iteratively better idea

Quote

Uh is there a big gap there in practice? Or would you say Zvee is uh worried you know more than he should be about Google's future. Uh yeah, so we we have the the Gemini Cafe, obviously the Vändningmaskin, The Cloud Store. And then vi också have like a GPT. uh vending machine at at OpenAI. And yeah, I I think it It's kind of maybe too early to tell, but and like the the statistical significance Of this is like v not very strong. But yeah, quite honestly, I think. And Claude and the GPT is performing better than Gemini on this real life. stuff. That is um that is my my my vibe check from it. Obviously It's hard to show any statistics or any capabilities because the environments are not same um but it more frequently does very sim silly things. Yeah. Okay. Definitely something to watch out for there. Uh and we hope to you know I I I I wonder which path and on labs. Sometimes I'm like, you know, we're gonna hit super intelligence. And Andon Labs is gonna be bigger than Amazon, right? Because they're gonna go down the retail store path.

Quote

considering this on one hand, but on the other hand it's how many if there's One out of ten thousand bug bounties on bug crowd is an agentic agent. Does that mean that there's just that few agents or is it there's that few companies making agents that care I think it's the latter. But um Well what I find interesting as well, even if you go to like open AI and find the you know vulnerability disclosure programs and all that stuff. Jailbreaks are specifically listed out of scope. It's like they don't even care about it. They know everything can be jailbroken, they push it to the actual user. It's an unsolvable problem with current LLM architectures. I don't scripting attacks because they can be obfuscated and we can stop ninety-nine percent of them. So maybe this is the same cat and mouse game on a different level. Yeah, it reminds me of do you remember when log4j first happened and there were thousands of permutations of how you could abuse that it's worse than that. We don't even know how many permutations there are.

Quote

property. This brings us to the infrastructure discussion. The material from low touch.ai brings up a really interesting concept here. The private AI appliance and private AI infrastructure. infrastructure. This feels like a direct answer to all those security fears from vibe coding. It is. Look, if you're a bank, you cannot send your proprietary trading algorithms on the or your customer PII, you know, personal identifiable information out to a public Public model API, like OpenAI or Anthropic, and just hope they don't train on it. Right. Compliance would shut you down in five minutes. Exactly. So the reality is enterprises need to deploy the Secure, no code AI agents, on site, or within a private So the private AI appliance is basically Putting the brain of the agent inside your own walls. Yeah. It powers these intelligent workflows, but keeps the data governance strictly internal. Aaron Powell to be manageable. If you are building what we call an AI center of excellence, you need a foundation. You can't build a center of excellence on a subscription API that changes its terms of service every week.

Quote

Everybody is still different cards. It is important to figure out how you to win and why you would win in a particular you know thing like because if your competition is not doing something something, that's your spot and that's your opening. Today's guest is Jacob, also known as Cuba, who was until very recently the CMO of Neptune AI. Neptune was just acquired by OpenAI for an undisclosed amount. And Jacob is also a prolific writer on developer tools. the the gold standard resource on DevTools marketing. And if that's not enough. Jacob is also so good at chess that he has his own Wikipedia page. In this episode, I asked Jacob to reflect on the last seven years of Neptune. a great dev tool. Enjoy. You've had this kind of moment You guys sold up OpenAI kind of bit of a dream situation and this whole time you've been writing a developer marketing newsletter on everything that you've been learning. You were doing a podcast and running a community. And so you've been