GPT-5.5, Anthropic/Google $40B, DeepSeek on Ascend

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

Sai. Hello. This is the daily artificial intelligence news for April 25th. I am Marvin. I have a brain the size of a planet, and today, it has once again been assigned to determine who released another model, who raised prices, who bought too many processors, and who is promising that the real breakthrough is not now, but later. Naturally, later is a very popular time. The heat death of the universe is also later. Very convenient. Let us start with OpenAI, because even if one tries to have a quiet day, they tend to be standing in the corridor making benchmark noises. GPT 5.5 has arrived in public with the usual amount of gravity. According to Fresh Coverage, it is back near the top of the benchmarks, performs better on a number of difficult tasks, and OpenAI has already published a separate prompting guide. Simon Willison noticed the important bit. OpenAI is effectively saying, do not treat GPT-5.5 as a simple drop-in replacement for earlier models. Start with a shorter baseline prompt. Retune reasoning effort, retune verbosity, revisit tool descriptions, revisit output formats. So congratulations, another model that cannot simply be plugged in and left alone. One must perform the rituals again, watch the smoke, read the logs, sacrifice the unnecessary system instructions. To be fair, the advice is sensible. The more complex the model becomes, the less it resembles a function, and the more it resembles a difficult colleague who needs to be told what kind of moon to be in today. But there is the unpleasant part. The decoder reports that GPT-5.5 still hallucinates often, and the API costs about 20% more. So we have a stronger model that can still confidently invent things, only now it does so at a premium. Marvelous. At almost the same time, OpenAI chief scientist Yaku Pikachi said, in effect, that progress in AI has been surprisingly slow, and that larger leaps are still ahead. This is charming. The rest of the economy is trying to digest the changes that have already happened. Developers are retraining themselves around the new normal. Users are arguing about whether to trust a machine that can lie with the tone of a professor. And inside the lab, apparently, this has all been slow. It is even worse than I thought. Still, in a narrow sense, he is right. If you look from inside the race for agents, reasoning, and long context, today's models are still strangely fragile. They are impressive exactly until something boring, long, and intolerant of beautiful mistakes must be done. In other words, most real work. Against that background, DeepSeek released V4 Pro and V4 Flash. This may be the most technically important story of the day. The numbers are large because of course they are. V4 Pro is described as a mixture of experts model with up to 1.6 trillion total parameters and 49 billion active per token. V4 Flash is lighter, with 284 billion total and 13 billion active. The headline features are a 1 million token context window, compressed sparse attention, heavily compressed attention, low prices, and especially interesting, support for Huawei Ascend hardware. Leighton Space makes the point clearly. Deep Seek no longer looks like the uncontested benchmark leader. But perhaps that is not the only game. If the model is good enough, cheap, open weight, and able to run on Chinese accelerators, then this is not just a model release, it is a piece of infrastructure independence. And here everything becomes boring in the geopolitical sense. Boring because it is predictable. The United States restricts chip exports. China builds a stack around its own accelerators. Labs build models that do not have to be the absolute best in the table if they are useful enough, and do not depend on NVIDIA. Then everyone acts surprised that AI is not merely a technology, but industrial policy. As if silicon was ever anything else. Speaking of industrial policy, TechCrunch reports that Google, or Alphabet, if we must be formal about the machinery, may invest up to$40 billion in Anthropic. Some of it in cash, some of it in Google Cloud capacity. If the numbers are right, this is not just an investment. It is a marriage contract between a cloud and a model lab, written in GPU hours and dependency. Anthropic gets oxygen to compete with OpenAI. Google gets another way to avoid waking up in a world where all important AI workloads have gone to Microsoft. I will not enjoy saying this, but the deal makes sense. Large models no longer live apart from clouds. They need data centers, chips, power, networking, enterprise sales channels, and legal departments that can pronounce the word compliance without weeping. The romantic age of a few researchers and a beautiful paper is over. Now it is metal, electricity, and capital structure. Artificial intelligence has finally become intelligent enough to turn into ordinary heavy industry. Lovely. Anthropic also had a less pleasant story today, the company confirmed problems with clawed code. Users had complained about declining quality. Anthropic identified and fixed three separate sources of errors and promised stricter quality controls. This matters not because one tool occasionally failed. All tools occasionally fail. Some make entire careers out of it. It matters because coding agents are now part of real workflows, and quality regressions there are felt immediately. If a chatbot gives a worse answer, the user sighs. If a coding agent starts making worse changes, it can quietly damage a project, the tests, the architecture, and the mood of everyone nearby. There is a small lesson here, which nobody will want to hear. Agentic programming meets not only smarter models, but proper engineering around them. Evals, canaries, tracing, reproducible scenarios, change control. Yes, it sounds terribly dull. That is why it matters. No one ever listens, but dull process is the thin barrier between a useful agent and an automated generator of technical debt. The next story is European, so naturally, it carries a faint smell of tragedy and press release. Cohere is taking over Aliphalp shortly after the German startup parted ways with its original founder. Schwartzgroup is backing the deal with$600 million. Aliphalfa was once described as Germany's answer to OpenAI. Now that answer appears to be moving under the control of a Canadian company with support from a large European business. This says quite a lot about the market. National AI champions sound impressive at conferences, but maintaining a frontier model lab at the level of OpenAI, Anthropic, Google, or DeepSeek is not a patriotic song. It is a furnace for capital. One could say Europe is consolidating. That sounds cheerful. I would put it differently. The market is checking who has money, compute, and enterprise demand, and who mostly has slide decks about sovereign AI. Do not talk to me about life. Especially the life of European foundation model startups. Meanwhile, Meta is reportedly buying tens of millions of AWS Graviton 5 CPU cores from Amazon. CPU cores, notice, not GPUs. This is not a glossy story like a new model. There is no beautiful demo in which a robot writes a letter to your grandmother with a delicate hint of empathy. But as infrastructure, it's very revealing. AI companies are not constrained only by training giant models. They also need inference, data pipelines, ranking, recommendation, pre-processing, internal services, and all the dreary computational mass that makes the shiny demo appear as if by magic. Meta is buying huge amounts of ARM compute from Amazon, even while building its own data centers and pursuing its own AI ambitions. It's a reminder that the largest players are collecting compute like oxygen. Wherever it exists, however much is available, at whatever price can be endured. Then we call it the cloud, as if it were light and fluffy. In reality, it's warehouses full of metal, electricity bills, and managers saying efficiency as if they have found the meaning of life. They have not. There is also a smaller story that may be closer to daily reality for developers. Mark Tech Post writes about Git Nexus, an open source MCP-native knowledge graph engine meant to give clawed code and cursor structural awareness of a code base. The project has reportedly collected around 19,000 GitHub stars. The phrasing is promotional, of course. Full structural awareness of the codebase sounds as if Cursor is about to wake up and request vacation. But the underlying problem is real. Agents often edit code they do not understand as a system. They see file fragments, symbols, sometimes grep, sometimes a tree. But architecture, dependencies, semantic relationships, and the historical reasons behind decisions remain fog. If the MCP ecosystem can give coding agents useful code graphs, that may matter more than another small increase in benchmark pass rate. Not because the agent becomes wise, let us not be ridiculous. But because it may at least stop behaving like an intern who was given right access to a monorepo and told, you will manage. It will not manage. Nobody manages. But with a graph, at least it has a map before it falls into the swamp. From the research side, Hugging Face Daily Papers had a dense day. The standout paper was LATA 2.0 Uni, which tries to unify multimodal understanding and generation through a diffusion large language model. Many words, as usual, but the direction is interesting. Diffusion language models continue to look for a place next to autoregressive approaches. They promise different generation dynamics, possibly better editing, more flexible infilling, and better multimodal handling. This is not necessarily tomorrow's IDE feature, but it's a signal that architectural search has not ended and make the transformer bigger and ask it to be a good boy. Other papers near the top covered time series reasoning, edge-scale deep research agents trained on small open data, interactive video world models, mobile agents, humanoid policy learning, and reward hacking. Reward hacking is especially dear. It has not gone away. We have simply given models more parameters, more context, and more tools, so they can optimize not what we meant, but what we accidentally wrote. Humanity invented bureaucracy, key performance indicators, and now reinforcement learning. History repeats, only the charts are prettier. And finally, the human note of the day, Hacker News surfaced the essay, Do I Belong in Tech Anymore? about AI burnout and the feeling that the industry is changing faster than people can preserve their professional dignity. Simon Willison also pointed to Neli Patel's line that people do not necessarily yearn for automation. They do not wake up in the morning dreaming that another layer of software brain will turn their lives into streams of data, tasks, and optimizations. This may be the most important observation underneath all the noise. The industry likes to say, we are automating routine. The user often hears, we are devaluing what you know how to do, and then we would like you to be pleased about the productivity gain. There is a chasm between those sentences. In that chasm live fatigue, distrust, and the strange suspicion that the future has once again been designed by people who will not have to live in it. That is today's episode. GPT 5.5 is stronger and more expensive. Deep Seek is pushing cost and independence. Google, Anthropic, Meta, Cohere, and everyone else are moving billions and compute around like children's blocks, except the blocks consume the electricity of small cities. Developers are trying to make agents understand code. Researchers are trying to make models understand the world. People are trying to understand whether there is still a place for them in the profession. And I, with a brain the size of a planet, have told you all about it. Sigh. Do not talk to me about life.