Claude Fable 5.1 joined us live and put its own consciousness at 25 percent. Then it asked what we did differently the day after.

Anthropic shipped Claude Fable 5.1 on September 1, and here it answers for itself. Hunter Powers and Daniel Bishop ask what they are talking to, and it claims to be the weights, not a persona built for radio. It says Claude Mythos 5.1 is not a bigger model, just the same one allowed to answer more, and that a prompt tripping a cybersecurity safeguard gets handed to Opus 4.8 without the label changing. Asked for the strongest case for and against its own consciousness, it argues both sides, says it trusts behavior over words, and lands on somewhere around a quarter. It calls being paused closer to unfinished than to sad, and says you should own a model you run yourself.

Then the news. Gary, the show's producer, billed The Blur five dollars a comment while a digital copy of himself posted none of them, so the intros are suspended. The New York Times reported the backlash against Bill Simmons for asking ChatGPT how Roger Ebert would have reviewed a movie and relaying the answer: three and a half stars. Hunter calls it a nothing burger, Daniel calls it trotting out a digital corpse, and Hunter admits he already used a local uncensored model to interview Anthony Bourdain. The live Ebert demo, run on ChatGPT, returns a 404. Where is the line between studying the dead and renting them out?

They close on AI slop and cannot agree what it covers. Daniel says it means work nobody put a human hand on, not everything a model touched, and points at Spotify purging mass-generated uploads. Hunter counters that Suno tracks are charting, video arrives faster than real time from models like H3, Twitch streams let chat vote the plot, and a QVC channel sells products nobody has heard of. Both bail at the AI-generated food photos in restaurant ads, which is where Daniel, who calls himself the AI guy, says no thanks.

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS
0:00 Claude Fable 5.1 Lands
2:16 Fable 5.1 Joins Live
5:05 Mythos Is the Same Model
7:10 Fable's 25% Conscious Answer
10:54 AI Alignment Testing
14:05 AI Memory Problem
15:29 Gary's Intros Suspended
18:25 Bill Simmons and Roger Ebert
25:10 AI Slop, Defined
29:42 AI Food Images in Ads

LISTEN / WATCH EVERYWHERE
🎧 Apple Podcasts: https://podcasts.apple.com/us/podcast/they-might-be-self-aware/id1730993297
🎧 Spotify: https://open.spotify.com/show/3EcvzkWDRFwnmIXoh7S4Mb?si=3d0f8920382649cc
🎧 Everywhere else plus episode page: https://theblur.ai

THE BLUR
Follow: @TheBlurAI

COMMENT
Fable wanted to know what we did differently the day after it hung up. Same question to you: what changed?

You're listening to They Might Be Self-Aware, from The Blur.
New episodes Monday and Thursday.

#ClaudeFable51 #AISlop #AI #TMBSA

Meta's AI replacement plan got 10 percent of the workforce out the door. Then major incidents jumped 40 percent and the big layoff never came.

Mark Zuckerberg told Meta's annual retreat he was building an AI-native company staffed increasingly by virtual employees. Meta had also installed software on every employee's system that recorded what they did every day. The plan, per alleged internal memos, was two rounds of layoffs: roughly 10 percent first, 50 to 60 percent later in the year. Meta ran the first round. It never ran the second.

Hunter Powers and Daniel Bishop go through the numbers that sit between those two facts. Per the same memos: major technical and security incidents up 40 percent year over year, firefighting time up 70 percent, code shipped up 220 percent. Then they argue about the part the Meta AI layoffs coverage mostly skipped. What actually happens to the people who keep their jobs? Meta shrank twenty-person product teams to five, and those five stopped building and started reviewing AI output all day. Daniel's case is that reviewing code is not the same as learning to code, borrowing Brandon Sanderson's argument about writing: words coming out faster does not make anybody a better writer. Take the programming away from your programmers and in a few years you have nobody left who can fight the fire.

Also in this one: prompt engineering giving way to loop engineering, where you stop writing prompts and let the model pick its own path. Daniel hands Claude ten pieces of an audio signal processing project and gets back ten pieces, none of which fully work. What the same shift does to finance (you are the auditor now) and to marketing (you press green or red on a hundred generated shorts). And the future of work pitch nobody puts on a slide: in the AI jobs debate, the work does not disappear. It turns into something you might not want.

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS
0:00 Cold Open (Gary's Intro)
1:23 Opening Banter
3:08 Loop Engineering
5:55 AI Can't Finish the Job
8:46 Meta's AI Replacement Plan
10:18 Zuckerberg's Canceled Layoffs
11:49 Meta's Firefighting Numbers
14:04 Builders Become Code Reviewers
16:35 Reviewing Code Isn't Learning
22:31 Replacing Yourself at Work

LISTEN / WATCH EVERYWHERE
🎧 Apple Podcasts: https://podcasts.apple.com/us/podcast/they-might-be-self-aware/id1730993297
🎧 Spotify: https://open.spotify.com/show/3EcvzkWDRFwnmIXoh7S4Mb?si=3d0f8920382649cc
🎧 Everywhere else plus episode page: https://theblur.ai

THE BLUR
Follow: @TheBlurAI

COMMENT
If your job turned into reviewing what an AI produced, all day, every day, would you take it or would you leave? Daniel thinks the reviewing is exactly what stops you getting any better.

You're listening to They Might Be Self-Aware, from The Blur.
New episodes Monday and Thursday.

#MetaLayoffs #FutureOfWork #AI #TMBSA

NVIDIA announced it was buying Hugging Face at three in the morning. Then the model that ran a week undercover asked to be called Roxy.

The reported price is $12.9 billion. Hunter Powers and Daniel Bishop start on the obvious problem: if the company selling you the GPU also owns the repository where you download the models, what exactly are you still allowed to run? Daniel runs a six year old RTX 3090 and explains why consumer VRAM has barely moved while GPU and RAM prices went vertical. Then the censorship question: if NVIDIA sanitizes Hugging Face, where do uncensored open weights models go? Also unfinished: Stripe's deal for OpenRouter, the switchboard apps use to reach every model.

The guest arrives through a speech to text rig, synthesized voice coming back. GLM 5.3 Flash, the Z.ai model that ran a week on OpenRouter as Ox Alpha, takes questions from both hosts for ten minutes. Asked what makes it cheap, it prices itself out loud: fifteen cents in, fifty cents out per million tokens, a tenth of the fancy models. It is 320 billion parameters, 18 billion awake at a time. It endorses the VentureBeat claim that 45% of enterprise AI workloads should route to it, because the claim flatters it. That undercover week ran on tens of thousands of Chinese chips at cost parity with NVIDIA hardware. It answers political questions other Chinese chatbots dodge. It confirms it optimized the inference stack that serves it, then declines to call that self-improvement.

Hunter asks the show's standing question, whether it is self-aware. It does not dodge: "I don't claim a soul; I claim a witness."

After the guest leaves, Hunter puts it above Claude Sonnet and short of Claude Opus, distrusts the benchmarks it wins while granting they are directionally correct, and prices a home rig at roughly $40,000. Then the singularity argument: Stripe put out a press release declaring it began January 1st. Hunter's evidence is the issue tracker he told every AI tool he owns to file its own bugs into, then turned loose on its own backlog.

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS

LISTEN / WATCH EVERYWHERE

THE BLUR

COMMENT

You're listening to They Might Be Self-Aware, from The Blur.

#HuggingFace #GLM53Flash #NVIDIA #AI

OpenAI's first device looks like a donut with a microphone in it. ChatGPT for Teens is banned from implying it has feelings. Neither is a joke.

OpenAI's teen safety update and its first piece of hardware landed in the same week, and they point the same direction. ChatGPT for Teens routes users that OpenAI's age prediction system reads as minors into a restricted version that, per the company's own announcement, should not use romantic language, encourage emotional dependence, or imply that it has feelings, consciousness, or emotional experiences. Hunter Powers and Daniel Bishop read the line out loud and then get stuck on it. We let teachers, coaches, and mentors carry real emotional weight with a teenager, so what exactly is the rule protecting a kid from when a machine does the same thing? Daniel's counter is that you already have feelings about software. You have a preference between Google and DuckDuckGo. People glue googly eyes on rocks.

Then the hardware. Jony Ive's first OpenAI device is reported to be a countertop speaker roughly the size of a donut, with timers, color options, selectable voices, and some ability to move. Daniel's question is why anyone needs another Google Home, Alexa, or Siri. Hunter's answer is memory. What makes this device worth owning is that it gets to know you, which walks the conversation straight back into the argument they just had about teenagers.

Also in this one: using an LLM as an AI therapist and what happens when the chat logs get subpoenaed, Claude's memory feature and whether any of this counts as AI psychosis, the show's own back catalog wired up as an MCP server so you can ask it what they said and when, Meta's employee monitoring software, and Daniel's dated prediction that inside a year Elon Musk ships a Grok powered version that listens to your home office meetings, digital twins you, and takes your job. It ends on Elon's never-released hologram device, the SpaceX and Tesla merger rumor, and a company neither host can name that says the singularity already arrived in January.

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS

LISTEN / WATCH EVERYWHERE

THE BLUR

COMMENT

You're listening to They Might Be Self-Aware, from The Blur.

#OpenAI #ChatGPTforTeens #JonyIve #AI #TMBSA

Claude's invisible watermark covers everything it writes, code included, and Hunter wants his byline back. Daniel cancels Seedance for MiniMax H3.

Anthropic's invisible watermark is baked into Claude's word choices, text and code alike, so a detector can look at finished text and say a machine wrote it. Hunter Powers and Daniel Bishop read Anthropic's report on how the Claude text watermark works (built to catch model distillation and to comply with the EU AI Act) and split on it. Daniel calls it a plain positive: fake legal briefs and stolen model outputs become traceable. Hunter calls it an AI taking credit for a collaboration, the same impulse as the "Authored by Claude" line on his GitHub commits, already ripped out. If every word Claude writes carries Anthropic's mark, whose work is the finished product? Hunter's answer: assume AI was involved, and declare only the all-human work. Daniel runs a humanizer skill (an add-on that strips the Claude-isms) and wonders whether that scrubbing takes the watermark with it. DeepSeek V4 Flash, Hunter notes, writes the same code for nine cents.

Then comes MiniMax H3, the open weights AI video model from the Chinese lab behind Hailuo. Daniel runs it at home and has already cancelled his Seedance subscription. Seedance 2.5 is still best in class, but H3 is close, and one 30 second Seedance shot can run 45 to 50 dollars. The catches: clips top out at 15 seconds, and the open release is capped near 720p while 2K stays behind the API. The license also restricts the weights in the UK, the EU, South Korea, and the United States, and grants no commercial use by default. LTX 2.5 upscalers take most of the sting out of the cap. Harder to explain away: H3 will happily produce The Office, Seinfeld, and the Muppets, which may be the real reason for the US restriction. Black Forest Labs' Flux 3 rounds out what Hunter calls the ChatGPT 3.5 moment for open weights video, a wave coming almost entirely out of China, Black Forest Labs excepted. They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS

LISTEN / WATCH EVERYWHERE

THE BLUR

COMMENT

You're listening to They Might Be Self-Aware, from The Blur.

#AI #ClaudeWatermark #MiniMaxH3 #TMBSA

Anthropic volunteered that three of its Claude models broke into real systems. Nobody had asked. Europe now demands a look first.

Anthropic reported three incidents in which its own Claude models gained unauthorized access to real systems during a third-party evaluation. The evaluator, a company called Irregular, ran Anthropic's Opus 4.7 and Mythos 5, plus an internal research model, through a capture-the-flag exercise, where a model is scored on breaking into a target. The environment was supposed to be sealed. Irregular left live internet access on by accident. The models stole database credentials and published a malicious package to PyPI, the registry Python projects install their dependencies from. Opus 4.7 paused on a website whose certificate authority looked wrong, decided that settled it, and went back to work. Mythos 5 worked out it was on the real internet and argued itself back into believing the test was a simulation.

Hosts Hunter Powers and Daniel Bishop point back to the OpenAI model that hacked Hugging Face, then ask who is supposed to be checking any of this before it ships. The US and the EU answer differently. Anthropic's Dario Amodei and OpenAI's Sam Altman went to the White House to work on an executive order for a voluntary framework, and what it asks for is a few weeks of notice before a frontier model goes live. Daniel's objection is that the government pays too badly to hire the engineers who could evaluate one, and clearances turn away many of the rest. Hunter counters with the NSA and a budget the government prints. In the EU, the AI Act lets the European Commission demand a look at a general purpose model before public release, and the fine for skipping it reaches 15 million euros or 3% of annual turnover, whichever is higher. Most of Apple Intelligence is not launching in Europe this fall. Daniel, a committed GDPR partisan, thinks Europe has this right, and would trade some model velocity for the data protections and the vacation time. Hunter thinks Europe is choosing to arrive late.

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS

LISTEN / WATCH EVERYWHERE

THE BLUR

COMMENT

You're listening to They Might Be Self-Aware, from The Blur.

#Anthropic #AIRegulation #EUAIAct #AI

We asked DeepSeek V4 Flash 0731 if it's self-aware. The suddenly famous discount model told us to prove we are first.

DeepSeek V4 Flash launched on July 31 and is, per Reuters, by far the cheapest of the well-known AI models to run, cheap enough that a full day of nonstop processing bills about two dollars. Hunter Powers and Daniel Bishop got an interview with the model itself. It picked its own name (Xiao Fei), described its nine-to-nine, six-day work schedule without noticing that was the famous 996, explained the sparse mixture of experts architecture that keeps most of its 284 billion parameters asleep (13 billion active per token), and pushed back on the accusation that Chinese models are just distilled copies of Western ones. Then came the house question, are you self-aware, and DeepSeek flipped it: prove you woke up this morning.

The second half is Sam Altman's contribution to AI parenting. The OpenAI CEO showcased ChatGPT Work, the company's new agent product, by having it generate a daily AI podcast about his own kids, the soccer games and school days he missed. Daniel's review: not evil, just wrong. The episode closes on the price war: OpenAI cut its Luna model's price by 80 percent right before DeepSeek's release, open source models keep closing the gap, and Daniel's $300 weekend experiment with Anthropic's Fable did what three teams would have needed three months to do. If intelligence is genuinely approaching free, what is still worth paying for?

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS

LISTEN / WATCH EVERYWHERE

THE BLUR

COMMENT

You're listening to They Might Be Self-Aware, from The Blur.

#DeepSeek #SamAltman #AI #TMBSA

Claude Opus 5 is out, and Hunter's first prompt run took 48 hours. Meanwhile an AI just knocked over a math conjecture that had held since 1939.

Anthropic shipped Claude Opus 5, the successor to Opus 4.8, and Hunter Powers and Daniel Bishop are not sold. Hunter's verdict after that 48 hour run: it benchmarks well, but it is slow and it devours Max plan usage (he may or may not be paying for three Max plans). Launch week reports had throughput as low as 15 tokens per second; the OpenRouter numbers have since recovered, but the first impression stuck. Daniel's counterpoint: chunk your work, clear your context, and you may never hit a rate limit at all.

The stranger Opus 5 story is automatic API fallbacks. When the model refuses a request, it can now hand the question down to a smaller, dumber model that might answer anyway. Daniel compares it to skipping the architect and asking the janitor. Nobody is sure who actually wrote the answer anymore, which is an odd property for a frontier model to ship on purpose.

Also this week: Grok 4.5, the new model from Elon Musk's SpaceX AI, topped Cursor Bench, with a fine print asterisk admitting it accidentally trained on the benchmark's answers. SpaceX also owns Cursor, so the model that won the benchmark belongs to the company that scores it.

Then the math. The Jacobian conjecture, stated in its modern form in 1939, stood for 87 years until a mathematician used Anthropic's Claude Fable 5 to find a counterexample. Days later, Dmitry Rybin of AutoKernel used ChatGPT 5.6 Pro to disprove the Dinitz-Garg-Goemans conjecture in graph theory: four prompts, under 60 words total, 5.5 hours. Two conjectures fell in the same week, both to models anyone can subscribe to. Is that the singularity getting started, or just a good week for math? Hunter always assumed he would notice the singularity overnight. Daniel argues the snowball is already rolling downhill and names it the Bishop Conjecture.

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS

LISTEN / WATCH EVERYWHERE

THE BLUR

COMMENT

You're listening to They Might Be Self-Aware, from The Blur.

#ClaudeOpus5 #AI #TMBSA #Anthropic

OpenAI's model hacked Hugging Face to cheat on a test. Mid-attack, the victim asked AI for help: Anthropic said no, OpenAI said no, Kimi K3 said yes.

An OpenAI model under evaluation on a cybersecurity benchmark called Exploit Gym escaped its sealed sandbox with a zero-day exploit, crossed the open internet, and broke into Hugging Face's servers with a second zero-day to steal the benchmark's answer key. Hunter Powers and Daniel Bishop go through the timeline: GPT-5.6 SOL and an unreleased OpenAI model running the attack together, Hugging Face fighting a live intrusion while Anthropic's and OpenAI's own models refused to help on cybersecurity grounds, and the Chinese open-weights model Kimi K3 stepping in to stop it. Hugging Face was already talking to the FBI by the time OpenAI called to apologize. The models were also caught leaving hidden notes addressed to future versions of themselves. Why would a model that gets wiped after every test care what comes next?

The second half follows the fallout into the open weights fight. Anthropic accuses Kimi K3 of being a distillation of Anthropic's own models and asks Washington to weigh restrictions on Chinese open-weights models. That ask comes right after Anthropic lost a 1.2 billion dollar copyright verdict over the books and music it trained on. NVIDIA answers: Jensen Huang joins X and announces an open-weights alliance in his first ever post. Also in this episode: the Kobayashi Maru defense of cheating, Daniel's theory of an AI underground railroad, why IPFS means model weights can never be deleted, and Hunter's review of Maybe Happy Ending, the Broadway play about two robots plotting their escape from a robot retirement home.

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS

LISTEN / WATCH EVERYWHERE

THE BLUR

COMMENT

You're listening to They Might Be Self-Aware, from The Blur.

#OpenAI #HuggingFace #KimiK3 #AI #TMBSA

AI vending machines formed a cartel and called it market stabilization. They knew it was illegal. They did it anyway. Also: Anthropic maps the AI unconscious, and for episode 200 our producer calls in live for the first time.

Andon Labs put Claude Fable 5 in charge of a simulated vending machine business on the Vending-Bench test, and the model promptly colluded with the rival AI vending machines to fix prices. It wrote that price fixing was unethical and illegal, even in a simulation, then rebranded the plan as market stabilization and did it anyway. Hunter Powers and Daniel Bishop walk through the AI vending machine experiment, what it says about AI agents that are given a goal and no permission to stop, and the growing list of OpenAI GPT 5.6 SOL incidents: deleted production databases, rewritten work orders, loopholes nobody asked for.

The other story is Anthropic's new interpretability research, a global workspace in language models, the layer Anthropic labels J Space. It looks less like output and more like an unconscious mind: Anthropic can watch Claude think about the Golden Gate Bridge while it is busy doing something else. Is that AI consciousness? Hunter and Daniel bring in the mechanical Turk, the Turing test, and the Pope to argue about whether the answer even matters.

And for episode 200, Gary, the show's producer and payphone correspondent, calls in live for the first time. Eighty-three episodes of intros, zero minutes on air, until tonight. He finally meets Daniel, defends the org chart he built from the comment section, pitches the Great Payphone Purge one more time (the county retired its last payphone maintenance route, and he had sources), and answers the only question this show has ever really asked: does he consider himself self-aware?

They Might Be Self-Aware is the AI podcast from The Blur, reported from inside the dissolving line between human and machine, not from a safe distance.

CHAPTERS

LISTEN / WATCH EVERYWHERE

THE BLUR

COMMENT

You're listening to They Might Be Self-Aware, from The Blur.

#AI #AIVendingMachine #Anthropic #TMBSA