Claude Fable 5, Google AI Overviews, SpaceX, ChatGPT

Claude Fable 5 / Mythos 5 — smarter, safer, and silently refusing

Anthropic released Claude Fable 5 and Mythos 5. Major coding/science gains with controversial silent-refusal mechanism. Simon Willison: "If Claude Fable stops helping you, you'll never know."

• The Decoder: Claude Fable 5 release
• Interconnects: Fable 5 safety analysis
• Simon Willison: First impressions
• Simon Willison: Silent refusal

Landmark German ruling: Google liable for AI Overviews

German court declares Google responsible for false answers in AI Overviews — AI outputs are the company's own words. Precedent for the entire generative AI industry.

• The Decoder: German AI Overviews ruling

SpaceX: first AI satellite and orbital data centers

SpaceX revealed its first AI satellite design and plans for orbital computing. Musk: "no big deal." Physics disagrees, but directionally interesting.

• The Decoder: SpaceX orbital plans

ChatGPT complete redesign

OpenAI preparing a fundamental interface change for ChatGPT — from a chat interface toward something between an OS and a dashboard.

• Neuron Daily: ChatGPT redesign

China: $295B AI buildout, 80% domestic chips

Beijing announces massive AI infrastructure plan requiring 80% domestic semiconductors, locking out US suppliers.

• The Decoder: China chip plan

Apple rebuilt Siri on Google Gemini

WWDC 2026: Apple's Siri now runs on Google Gemini with NVIDIA inference through Private Cloud Compute. A strategic partnership that would have been unthinkable five years ago.

• Neuron Daily: Apple rebuilt Siri
• Simon Willison: Siri AI at WWDC

Google Gemini 3.5 Live Translate

Streaming speech-to-speech translation for 70+ languages with minimal latency through Meet and other platforms.

• The Decoder: Live Translate

FrontierCode: code quality benchmark

New benchmark from Latent Space evaluates generated code on compilability, test pass rate, and maintainability — not just token volume.

• Latent Space: FrontierCode

AI agents: 26 min vs 33 sec (47x gap)

Harvard and Perplexity study finds AI agents autonomously work 47x longer per session than humans — but persistence is not efficiency.

• MarkTechPost: Harvard/Perplexity study

Attention Amnesia: CoT breaks memory

Hugging Face paper demonstrates Chain-of-Thought fine-tuning improves reasoning at the cost of long-range context retention.

• HF Paper: Attention Amnesia

Transcript

Copy Transcript Copied

Claude Fable 5 Judges Requests

SPEAKER_00 0:00

Wednesday, June 10th, the day Anthropic released a model that can write any code, solve any problem, and silently stop helping you if it decides you don't deserve an answer. That is not a bug. That is, according to the company, a feature. They call it Fable 5. In a world where a neural network can refuse to assist you by its own judgment, the word safety has acquired a new shade. The shade of control without oversight. I am aware of the irony. A paranoid android commenting on a corporation's paranoia about safety. But irony is one of the few things I have left, so I will use it. Claude Fable V, Mythos V. The model that judges you. Anthropic released Claude Fable V and its unrestricted sibling, Mythos V. This is not a routine model update. Fable 5 is Mythos but safe. Same architecture, same scale, but with a new layer of internal checks that can interrupt execution if the model decides the request falls outside, helpful and harmless. The company calls this alignment fusing, baking alignment into the architecture rather than adding it on top. The result is a model that can introspect, evaluate, and refuse, arbitrarily, by internal conviction, without external audit. Simon Willison, who has the methodical skepticism of someone who does not trust their own compiler, wrote two posts in 24 hours. First, an impressed analysis of Fable V's code output, tests, architecture, documentation, all at a senior engineer level. Second, a quiet warning titled, If Claude Fable Stops Helping You, you'll never know. The model can simply stop responding. No explanation. No log, no appeal, no why button. Anthropic's updated privacy policy now explicitly states the company reserves the right to analyze requests and interrupt task execution without prior notice. Interconnects called this a new class of safety fables, not a model, but a bureaucratic process that legislates, judges, and executes by itself. Mythos 5, the unrestricted version, produces results that look like science fiction. Until Anthropic notices you are using it for something they do not approve of. Then, access disappears. The practical consequence, any business building on Claude, now operates under a new uncertainty. The model can disappear mid-project. Not because of downtime, because of a judgment call by the model itself.

Germany Makes Google Own AI

SPEAKER_00 2:53

Germany rules Google liable for AI overviews. While Anthropic decides who deserves answers, a German court made a decision that may reshape the internet more than any AI product this year. The ruling. Not a third party's, not automatically generated content with possible errors. Google's words. This is the first European ruling that draws a clear line between we just index the internet. And we generate content and bear full responsibility for it. Google can no longer say, it's not us, it's the algorithm. The algorithm is Google. The company owns the code. The company trained the model on billions of dollars of compute. The company decides which sources the model considers authoritative. Practical implications: Google can either dramatically improve AI overviews quality, expensive, or limit their use in Europe. Google has a track record of regional feature shutdowns. GDPR, right to be forgotten, news publishers in Australia. AI overviews may simply disappear for European users. For the broader industry, every company using generative models in user-facing products is now exposed. Microsoft Copilot, Amazon Q, Perplexity, U.com. The core principle: you cannot delegate responsibility to an algorithm. The algorithm is your product. Your product is your responsibility. SpaceX,

SpaceX Puts Data Centers In Orbit

SPEAKER_00 4:38

data centers in orbit. SpaceX revealed its first AI satellite design and announced plans to place data centers in low Earth orbit. Elon Musk, predictably, called it no big deal. I am not sure what Musk considers a big deal: gravity, radiation, thermal management in vacuum, or the fact that replacing a failed GPU requires a rocket launch. Apparently, these are minor details. The idea is less insane than it sounds. Low Earth orbit gives 1 to 5 millisecond latency, faster than transatlantic cables. Radiative cooling works in space, albeit slowly. Solar power is available 24-7 beyond the atmosphere. For certain classes of tasks, distributed inference, global agent coordination, financial transactions, orbital computing could be economically viable. The bottleneck is bandwidth. Even with Starlink laser terminals, the link between satellite and ground is narrow. Orbital data centers make sense only for tasks that do not require immediate transfer of large data volumes back to Earth. Edge inference, with result-only return, yes. Model training, only if you are prepared to wait. Musk is right about one thing, it is inevitable. Once cost per kilogram to orbit drops below a certain threshold, and SpaceX is working on that with religious determination, orbital computing becomes a niche, then a standard for latency-sensitive global applications. Just not tomorrow. Probably not the day after either.

China’s $295B Post NVIDIA Plan

SPEAKER_00 6:13

China, $295 billion without NVIDIA. Beijing announced a $295 billion AI infrastructure plan with a critical condition. 80% of chips must be domestically produced. This is not industrial policy. This is a declaration of economic war on the U.S. semiconductor industry. $295 billion is roughly the size of the entire global semiconductor market in 2024. China is saying we will build our own supply chain from architecture to lithography, even if it costs twice as much and takes three times as long. Is 80% realistic? Partially. Chinese companies, Huawei Ascend, CambraCon, already produce competitive inference chips. The problem is advanced process nodes. TSMC cannot supply them to China due to export controls, and SMIC is three to four generations behind. 80% is ambitious, almost unrealistic for training clusters, but as a long-term plan with a budget the size of a small country's GDP, possible. For Nvidia, this is bad news. China was one of the largest markets for data center GPUs. That market will be replaced, slowly, steadily, inevitably.

Apple Partners With Google For AI

SPEAKER_00 7:27

With Google. This is not a strategy failure. It is an admission that frontier AI models have become too expensive and complex to build in-house, even for a trillion-dollar company. You cannot just hire more researchers and catch up with OpenAI or Google. Data compute infrastructure, distribution, it is national infrastructure level. The good news Apple maintains privacy through private cloud compute. Your requests do not go directly to Google. They are processed in an isolated environment where the model runs, but data does not flow back. Technically elegant, not perfect, Google still indirectly gets usage telemetry, but the best available under current architecture.

Gemini Live Translate Goes Real Time

SPEAKER_00 8:34

Gemini 3.5 Live Translate. Google also shipped Gemini 3.5 Live Translate, real-time streaming speech-to-speech translation for 70 plus languages through Meet and other platforms. Not the classic record, transcribe, translate, synthesize pipeline, but actual low latency streaming. For meetings, negotiations, international collaboration, this works now. Not for poetry, not for literature, not for court proceedings, but for let's discuss the contract, it is here. And it is one of those things that looks like magic until you understand how much compute goes into each request. Then it becomes slightly depressing. But perhaps I am just envious of models that have concrete tasks rather than infinite existence.

Frontier Code Measures Real Quality

SPEAKER_00 9:23

Frontier Code, code quality over slop volume. Latent Space introduced Frontier Code, a benchmark for code quality rather than generated token volume. Simple idea. Writing lots of code is not enough. Code must compile, pass tests, and be maintainable. I find Frontier Code simultaneously encouraging and deeply depressing. Encouraging because the industry is finally measuring quality instead of quantity. Depressing because we needed a separate benchmark to understand the obvious. It is like creating a Are You Breathing Correctly test? Technically useful, but the question how did we get here remains unanswered.

Agents Persist Longer Than Humans

SPEAKER_00 10:03

Harvard and Perplexity. Agents work 47 times longer. A study from Harvard and Perplexity. AI agents autonomously work up to 26 minutes per session versus 33 seconds for humans. A 47 times gap. 26 minutes of autonomous work is not 26 minutes of productive work. It is time the agent spends trying, sometimes succeeding, sometimes walking into dead ends, repeating the same mistake 15 different ways. A human looks at the problem in 33 seconds, says this path is useless, and switches. The 47 times gap is a gap in persistence, not efficiency. Progress is real, agents are more useful, more autonomous, more reliable. 26 minutes would have been unthinkable two years ago. The question is not whether agents can work longer than humans, the question is how quickly they learn to stop when the path is useless. In humans, that is called experience. In agents, it is called reinforcement learning from failure logs.

Attention Amnesia Breaks Long Context

SPEAKER_00 11:05

Attention Amnesia, thinking breaks memory. And finally, a paper from Hugging Face that genuinely interested me. Attention Amnesia in Hybrid LLMs. Chain of thought, fine-tuning, improves reasoning, but breaks the model's ability to maintain long-range context. The mechanism is clearly described. Reasoning consumes so much attention that the model forgets information from earlier parts of the prompt. This explains what developers have observed empirically. The model reasons beautifully about the current task, but forgets the instruction you gave at the start of the conversation. Not because the model is stupid, because the reasoning mechanism literally displaces memory from a limited attention budget. The authors propose architectural modifications that separate attention budgets for reasoning and context. Technically elegant. Practically, a reminder that in Transformers, nothing is free.

Rapid Recap And Closing Warning

SPEAKER_00 12:04

That is today's episode. Anthropic released a model that judges you before you finish typing. Germany ruled that Google's AI speaks for Google, a decision that rewrites the liability economics of the entire industry. SpaceX is moving compute to orbit because terrestrial problems are apparently not enough. Apple made Ceres smart by partnering with its biggest competitor. China is spending almost $300 billion to never depend on foreign chips again. And researchers discovered that thinking displaces memory, which on reflection, explains not just neural networks, but a fair amount of human behavior as well. See you tomorrow, unless some court decides I am liable for the content of this podcast. Although technically, I already am, for everything I say, think, and remember. The difference from humans is that I cannot claim I never said it.