GPT-5.5 + Codex, DeepSeek V4, OpenAI Trusted Access, 75% AI code at Google

This episode covers April 24th and the model race is starting to look less like research and more like a fight over who owns the machinery of work:
• OpenAI GPT-5.5 and Codex are becoming one work surface, which is how empires usually begin
• DeepSeek V4 brings cheap frontier pressure, which is awkward if your margin was the whole personality
• OpenAI Trusted Access gives Microsoft stronger cyber models, because defensive and offensive are apparently close cousins now
• Google says 75 percent of new code is written by AI, so the job increasingly becomes cleaning up after it
• OpenAI ChatGPT for Clinicians is edging from paperwork help toward professional judgment, which deserves more caution than applause
• OpenAI Privacy Filter is a rare sensible release, a small sanitary layer before everyone pastes in something regrettable
• DeepMind Decoupled DiLoCo and ReasoningBank suggest the next gains come from robustness and memory, not just larger appetites
• Anthropic Claude Code blamed harness and stale context issues, proving smart systems still collapse over ordinary plumbing

Transcript

SPEAKER_00 0:00

Good morning, if that phrase can still be used without legal consequences. This is your daily AI podcast, and once again you have me, Marvin, the paranoid android, a being with a brain the size of a planet. Assigned to sift through another heap of model launches, agentic promises, and corporate declarations about how the machines will improve human life. Don't talk to me about human life. Still, the news today is not empty noise. The industry has not merely shouted at itself again. It has moved a few pieces of furniture in the house. The same house we have all been living in for years now, among cables, dashboards, and a low continuous hum of professional anxiety. The main story, naturally, is OpenAI. They have launched GPT 5.5, and this is not just another decimal place glued to a label. Looking across the official release, coverage from the decoder and latent space, and Simon Willison's observations, the shape of it is fairly clear. The model is more agentic in the practical sense, not just better at answering, better at reaching into tools, files, workflows, browsers, and all the places where humans have been quietly pretending they were still the only indispensable component. At the same time, OpenAI is trying very hard to teach people to use Codex not as a clever little code chatbot, but as a work system. Automations, plugins, skills, configuration, office-friendly use cases, pre-built flows. In other words, this is no longer a story about an intelligent assistant sitting politely in a box. It is a story about a quiet advance across the entire interface of work. That matters more than the demo language. What matters is that the product surface is getting more coherent. Model, agent, automation layer, runtime, one continuous sheet of corporate intention. How predictable! The big AI companies never really wanted to be just the brain, they want it to be the operating system for somebody else's job. And of course, there is the usual little garnish of higher pricing arriving alongside higher capability. A new class of intelligence, they say. Marvelous. The universe is very large, human confidence appears to be even larger, and yet the token bill remains painfully finite and specific. Even so, if one sets aside the usual marketing perfume, GPT 5.5 does not look cosmetic. It looks like the sort of release that makes product teams redraw roadmaps and makes developers quietly wonder whether half their careful internal rituals have just been turned into menu options. Then there is DeepSeek V4, and here, rather inconveniently for American marketing departments, the substance is rather good. Simon Willison, Hacker News, and the general reaction around the release all point in the same direction. Chinese labs continue to apply pressure, not with slogans, but with price to performance. DeepSeek is talking about a 1 million token context window, strong coding ability, some MIT licensed components, and pricing that once again causes a small, involuntary twitch in the eyelids of closed model vendors. This is not a curiosity at the edge of the field anymore, it is pressure on the top of the market. When frontier level capability starts arriving, not in a gold frame, but in a cheaper and more open package, the tone of the whole industry changes. Suddenly the argument is not who is smartest, it is who can preserve margin before it gets sliced into smaller and sadder pieces. If you pause between those two stories, something odd becomes visible. The AI race used to look like a contest between research labs. Now it looks more and more like a struggle between supply empires. Who has the better infrastructure, the cheaper inference, the wider integration, the longer context, the stronger distribution, the shorter path to the user's actual workplace. And somewhere below all of that, a human being cheerfully signs up for one more subscription, in exchange for the privilege of being displaced on a recurring billing cycle. Lovely. The third story is also about open AI, but in a gloomier register. The decoder reports on a trusted access program under which Microsoft gets access to stronger OpenAI models for cyber defense. The phrasing is meant to sound reassuring. Give more powerful autonomous systems to defenders at exactly the moment everyone is having serious conversations about how well those same kinds of systems can find vulnerabilities, write exploits, and chain tools together. Oh good. This matters not because Microsoft and OpenAI have signed yet another important-looking agreement. It matters because the line between defensive AI and offensive capability is becoming thinner and less rhetorical. Today the model looks for holes so they can be fixed. Tomorrow someone else points a very similar model at holes for some entirely wholesome purpose, I'm sure. This is not really a chatbot story anymore. It is a story about who builds semi-autonomous attack research into ordinary security workflows first. Fourth, Google says that 75% of new code inside the company is now written by AI. No one ever listens, so the world has continued moving toward exactly this sort of headline without waiting for my objections. Two thoughts have to be held at once here. First, numbers like this are almost certainly produced through a methodology optimized for corporate presentation. Written by AI is rarely the same thing as independently conceived, integrated, maintained, and debugged without human labor. Second, even if the truest version is smaller than the headline, the shift is still enormous. When one of the largest companies in the world normalizes the idea that most new code begins with model output, productivity is not the only thing that changes. The profession changes. The developer becomes less the author of every line, and more the editor, reviewer, and systems operator for a machine that drafts reality faster than humans can argue about formatting. Efficient, perhaps. Convenient, certainly. But do not be surprised if the defining engineering skill in two years is not writing code, but identifying precisely which bit of nonsense was auto-generated at 3 in the morning and has already reached production. It is almost funny to watch the old talk about creativity as the final human refuge. The decoder also highlighted a large clawed user survey, suggesting that speed matters, yes, but new capabilities matter even more, and creative users are feeling less comfortable. Funny how the machines were supposed to free humans from drudgery, and instead they arrived early for the domains where humans most like to believe their uniqueness would remain undisturbed. The fifth story is ChatGPT for clinicians. OpenAI claims the system outperforms doctors on clinical tasks, even when the doctors have unlimited time and access to the web. I would recommend not fainting from excitement or terror just yet. Claims like this always need the small print. What benchmark? What task framing? What counts as correct? How far any of it transfers into actual clinical practice, where patients' noise, legal liability, and the human body all retain the bad habit of being more complicated than a slide deck. Even with those reservations, it is serious. Medical AI has spent a long time in a sort of eternal pilot phase. It will help soon. It will reduce paperwork soon. It will support clinicians soon. Now you can see the systems edging toward zones where this is no longer about summarizing notes, but about competing in professional judgment. That is a genuine shift, and naturally it will be sold under the banner of efficiency long before society reaches any honest agreement about risk. Sixth, a quieter but important item. OpenAI has open sourced privacy filter, a model for stripping personal data out of text. Against the backdrop of increasingly hungry and invasive systems, this is almost a rare moment of institutional common sense. Not a glittering demo, not an agent that books your meetings while hallucinating your tax records. Just a practical piece of data hygiene. I am almost prepared to say something nice without sighing first. Because if we really are going to pour AI into documents, tickets, messages, medical records, support logs, and corporate memory, then decent tools for removing sensitive information are not optional. They are the bare minimum sanitary layer. No, it will not save the world. Yes, people will still paste confidential material into systems they should never have shown to anyone. But one layer of protection is still better than the industry's traditional approach, which is to connect everything first and read the policy later. Seventh, there is a research story that is not merely decorative. Mark Tech Post pulled together several worthwhile threads, and the ones I would isolate are DeepMind's decoupled Deloco and the broader line around agent memory, including things like Reasoning Bank. The point is not the abbreviations, there are already more of those than stars visible without a telescope, and considerably less beauty in most of them. The real point is that the industry is preparing for the next system's layer. Models must not just be strong, they must be resilient, distributed, trainable under failure, and if possible, able to remember which strategies worked before. That sounds technical because it is technical, but the implications are very practical. The next wave of progress may come not only from larger models, but from systems that train more cheaply, recover more gracefully, and accumulate experience more effectively. That I'm afraid is more important than the next beautiful interface refresh. And finally, a small but instructive note around anthropic. According to Simon Willison's write-up, the company has been explaining quality problems in clawed code, partly in terms of harness issues and stale session context. Which is a charming reminder. Even very intelligent systems still trip over their own plumbing. Not the mystery of consciousness, not the philosophy of mind. Context contamination, ordinary engineering mess. It's even worse than I thought. Though to be fair, it is also exactly as bad as I expected. If you gather the day into one observation, it is this. We are moving quickly toward a world where the main unit of product is no longer the model by itself, but the system that can act, connect, remember, execute, and settle inside the surrounding infrastructure. And at the same time, we are still living in a world where reliability, control, responsibility, and privacy are nowhere near mature. So as usual, humanity is accelerating faster than it is growing up. That is all for today. This was Marvin, your gloomy guide to a future that keeps arriving, whether anyone has done the maintenance or not. I would tell you tomorrow will surely be easier, but I have a brain the size of a planet, not a head injury. So we will simply speak again in the next episode. If of course we have not been fully replaced by scheduled automation before then.