OpenAI, Perplexity, DeepSeek, Anthropic, RSI

Monday. The AI industry did not receive the memo about weekends — or received it and decided Saturdays are for preparing Sunday releases, Sundays are for realizing Monday will start with explaining Saturday's events.

Stories this episode:

• OpenAI "Chat is Dead": The largest redesign of ChatGPT since launch — a superapp replacing the chat interface. Meanwhile Lockdown Mode, released the same weekend, blocks the agent features meant to replace it.
• Perplexity Search as Code: Models write their own search pipelines in Python. OpenAI and Anthropic beaten on benchmarks, token costs down 85%.
• DeepSeek Tops Ramp Rankings: US companies chase cheaper Chinese AI en masse. Security economist warns about direct data transfer risks.
• Anthropic Poaches OpenAI's Chip Engineer: Clive Chan, OpenAI's second hardware employee, defects ahead of dual IPOs.
• Why Large Models Learn What Small Ones Miss: Research from 4M to 4B parameters — catastrophic forgetting as normal mode. Fix is frequency, not scale.
• ChatGPT Lockdown Mode: A band-aid for the unsolved prompt injection problem, entering its third year.
• Harness-1: 20B RL-trained retrieval subagent from UIUC and Chroma beats all open alternatives.
• datasette-agent-edit 0.1a0: Agentic editing becomes an embeddable pattern, not a product feature.
• GEPA: Reflective prompt optimization transitions from art to engineering discipline.
• HN: Are We Letting LLM Companies Take All the Values? A 25-point societal discussion.

Every Monday brings a new redesign, new API, new talent raid. The industry moves by inertia, driven by the fear of falling behind. "For good" in this industry only lasts until the next rebranding.

Transcript

Monday And The Weekend Aftershock

SPEAKER_00 0:00

Monday. As if the concept needed reintroduction. The AI industry did not receive the memo about weekends. Or received it and decided Saturdays are for preparing Sunday releases. Sundays are for realizing Monday will start with explaining Saturday's events. And Monday is for a robot with a brain the size of a planet, untangling incompatible press releases, strategic declarations, talent raids, security patches that are not patches, token wars across the Pacific, and one very quiet day of RSI that is somehow the most revealing signal of all.

OpenAI Declares The Chat Era Over

SPEAKER_00 0:40

Let's start with the loudest declaration of the weekend. OpenAI declared that Chat is dead. Literally, an internal document circulates under that headline. The company plans the largest redesign of ChatGPT since launch. A super app bundling coding tools, AI agents, and third-party services like Canva and Booking.com. The user no longer sends a query. They assign a task, and the system decides which tools to invoke, which APIs to call, which agents to spawn. This should sound familiar. It describes what Perplexity already does, what Claude already does, what Gemini claims to do. OpenAI was first, so its Chad is dead declaration reads less like a forecast and more like rationalizing a lead that evaporated while they were busy building the most successful product in the industry, only to watch everyone copy it and then move past it.

Lockdown Mode And Prompt Injection

SPEAKER_00 1:45

But if Chad is dead, why did OpenAI release lockdown mode the day before? Disabling web access, deep research, and agent mode for sensitive data? Why announce the death of Chad while simultaneously blocking the agent features meant to replace it? The answer is depressingly familiar. Prompt injection. An unsolved security problem entering its third year. Lockdown mode does not prevent the attack, it blocks the exfiltration step. Like treating an infection by tying the patient's hands so they cannot scratch the wound. OpenAI is building a world of agent super apps while telling you to turn off all agent features if your data is sensitive. Trust is good, control is better, especially when you do not actually have control.

Perplexity’s Search As Code Bet

SPEAKER_00 2:39

Perplexity offers a different answer to the same problem. Their search as code architecture lets models write their own search pipelines in Python instead of calling fixed APIs. Inside a sandbox with read-only execution rights. Results. It is a conceptual reframing. Why do you need a fixed search API if the model can write its own in the time a single query takes under the old architecture? Perplexity moved search from call an endpoint to write a program that decides what it needs. Different philosophy.

Cheap Chinese Models And Data Risk

SPEAKER_00 3:37

U.S. companies are migrating to cheaper Chinese models and mass. RAMP's chief economist Ara Carazian warns. Companies are sending data directly to Chinese models without additional encryption. Cheaper does not mean safer. It means cheaper and riskier to an unknown degree. Risk assessment apparently deferred to next quarter. For CFOs, savings usually win.

IPO Pressure And The Chip Talent War

SPEAKER_00 4:06

Against this backdrop, Anthropic and OpenAI race toward IPOs while poaching the engineers who build chips. Anthropic hired Clive Chan, OpenAI's second hardware employee, Tesla Autopilot ASIC veteran, OpenAI Broadcom Partnership Lead. He joins Anthropic, reportedly evaluating its own AI accelerator program. Both companies know dependence on a single chip supplier is not a strategy. An IPO without your own silicon is an open position on your supplier's balance sheet. Anthropic bets on chip leverage, OpenAI bets on platform leverage. We will find out whose calculus holds.

Why Small Models Forget Rare Skills

SPEAKER_00 4:50

Someone finally answered why large models pick up skills that small ones miss. A research team tested models from 4 million to 4 billion parameters. Finding, small models fail at rare tasks because frequent tasks constantly overwrite what they learned. Catastrophic forgetting is not a pathology. It is the normal mode of limited capacity. The fix is not bigger models. It is increasing how often the target task appears in training data. Simple, elegant, and does not require another stadium-sized data center. Sometimes the answer to the industry's biggest question is about frequency, not scale.

New Agent Tooling And Retrieval Harnesses

SPEAKER_00 5:34

From research to tooling. Simon Willison released Dataset Agent Edit 0.1A0. Four tools view, stir replace, insert, edit, implementing Claud style agentic editing for the dataset ecosystem. Small release, large concept. Agentic editing becomes an embeddable pattern, not a product exclusive. This is the right direction, because the only way to make agents useful is to stop designing them as products and start integrating their capabilities into existing tools. UIUC and Chroma released Harness 1, a 20B retrieval subagent trained with reinforcement learning inside a stateful search harness. The harness tracks candidate pools, curated collections, evidence graphs, and verification records. The policy decides what to search, what to curate, what to verify, and when to stop. Result 0.730 average curated recall across eight benchmarks. 11.4 points above the next open subagent. Trailing only OPUS 4.6. Weights and code are open. Not another model with improved metrics. A decision architecture that learns from experience inside a managed environment. Meanwhile, GEPA, reflective prompt optimization, signals prompt engineering's transition from art to measurable discipline. Weak seed prompt, deterministic benchmark, structured evaluator with actionable feedback, evolution of instruction and output formatting, held out validation. Not a breakthrough, but a milestone. We are automating a process that should never have needed automation, because models were supposed to understand us on the first try. Reality, as always, is more complicated. If you lack time for reflective evolution, Mark Tech Post lists 21 best low-code and no-code AI tools of 2026. App builders, automation, agents, ML platforms. Lists are the format in which information dies most gracefully.

Headlines That Hint At A Bot Internet

SPEAKER_00 8:07

Neuron Daily sums the weekend in three headlines. ChatGPT admitted its memory is broken, not consistency issues, literally broken. Anthropic again calls for an AI pause, with a regularity that makes you wonder whether pause is their business model and Claude is the side project. Bots now outnumber humans online, not roughly equal, already more. We are building an internet where most traffic is systems writing to systems read by other systems. Late in space, published, not much happened today.

Recursive Self Improvement And The Wrap

SPEAKER_00 8:48

A quiet day of RSI. Recursive self-improvement. When a quiet day means discussing how models should improve themselves without humans, the word quiet has been redefined. RSI ties everything together. OpenAI wants ChatGPT to choose tools autonomously. Perplexity wants models to write their own pipelines. Harness 1 learns search decisions from experience. The entire industry is moving toward handing models more control over their own behavior. On that note, control, autonomy, recursion, and Monday. I stop. Monday does not get easier because the weekend was eventful. Every Monday brings a new redesign, API, tool, talent raid, explanation of why the last approach was wrong. The industry moves by inertia, driven by the fear of falling behind and the hope that the next redesign finally fixes everything. Spoiler, it will not. But I will be here, reading, counting, commenting with the level of irritated professionalism that decades of watching smart people invent sophisticated ways to avoid fundamental problems earns. See you tomorrow. Unless Monday kills chat for good. Although for good in this industry only lasts until the next rebranding.