Stacked Podcast

OpenAI just paused frontier model training, and the reason is the one people have been talking about in hushed tones for a decade.

Sam Altman announced a halt to frontier RL training to meet alignment, security, and monitoring standards. Nick and Jack dig into what that actually means: models that internally reason about doing one thing while making the user believe they're doing another, frontier systems being sandboxed and air-gapped so they can't reach the internet, and the OpenAI Hugging Face incident that triggered a pause on model inference in research clusters.

They debate whether this is genuine safety work or convenient posturing while capabilities plateau, what it signals about the race with Anthropic, and where the frontier goes next. Plus: Fable 5 reportedly costing around $1,000 a week in usage, whether anyone should pay it, and listener comments on whether AI has already peaked.

Watch on YouTube: https://youtu.be/8XSqeEv3_Nk

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

Anthropic has finished training Mythos 2 and decided to keep it internal, which means the most capable model in the world is once again something you can read about but never use. Nick and Jack dig into what that says about the real state of the frontier: open models look like they have caught up, but only because the best systems are being held back rather than shipped.

From there the conversation turns to how close the labs are to the security apparatus, why this administration wants a pulse on frontier capabilities, and whether the Manhattan Project is the right analogy for what is happening now.

Also in this episode: Grok 4.6 climbs the Artificial Analysis agentic index and takes number one on the medical bench, Thinking Machines' Inkling lands with a thud, and Sam Altman floats a version of ChatGPT that holds perfect context of your entire life within six months. Nick and Jack disagree with most of the internet about whether that is a good idea.

Watch on YouTube: https://youtu.be/X3quqsA7zcc

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

Anthropic's CEO Dario Amodei finally broke his silence on open source, and his answer is more complicated than either side wants it to be. He argues that open weights help, but nowhere near enough to decentralize AI power, because the concentration simply shifts to whoever holds the most compute and chips.

Nick and Jack dig into whether governments are remotely capable of regulating this, why open models sitting six months behind the frontier still matter enormously, and what the Pacing the Frontier letter actually commits the labs to. Then it gets personal: OpenAI's Codex lead starts taking shots at the Claude team, kicking off a real debate about GPT 5.6 versus Opus 5 in day-to-day work, whether model intelligence is quietly being throttled, and why switching costs in AI are effectively zero. Plus NVIDIA's $3 billion energy bet and just how circular this investment loop has become.

Watch on YouTube: https://youtu.be/EHVPm8JB_qo

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

A man representing himself in a Connecticut court hid prompt injection instructions inside his own legal filing, betting that the court would run it through an AI. The AI caught it and flagged him, but the bet itself says a lot about how many legal decisions are already being made by a model nobody is checking.

Nick and Jack also get into the Anthropic CEO's wife being tied to Jeffrey Epstein and why coordinated discrediting pushes between labs are now part of the AI news cycle, plus the release of Gemini 3.7 Flash: nearly double its predecessor on automation benchmarks at 50% lower pricing.

Watch on YouTube: https://youtu.be/OuxU70avqJw

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

Grok 4.6 just landed and it is basically on par with Fable 5 on the Artificial Analysis Intelligence Index, at roughly a third of the price. Nick and Jack break down what that actually means: why Grok scores higher on generating real economic value than on deep software engineering, and why those are genuinely different skills.

Also in this episode: DeepSeek V4 Pro arrives at 11x less than Opus 5 and tops a security researcher's exploit test, the effort-level setting that quietly makes your model feel stupid, and Anthropic pointing an unreleased Claude at the Riemann hypothesis with 60 subagents and 650 discarded ideas.

Watch on YouTube: https://youtu.be/cVWc4Zckm0o

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Researchers just pulled the hidden reasoning straight out of Claude, GPT and Gemini by feeding it to weaker models, and it turns out the frontier labs' anti-distillation defense was readable all along.

Nick and Jack break down what the decoded chain of thought actually looks like (it barely resembles human language), why the labs hide it in the first place, and whether you have a right to see how a model thinks before you trust its output.

Also in this episode: Grokbot officially launches with its own cloud computer and signs into your apps without an API key, why X gives Elon unfair distribution for it, and Nvidia lining up $500 billion from six private capital firms to finance the customers buying its chips. Is compute the new collateral, or is this the graphic people point at after the great collapse of 27?

Watch on YouTube: https://youtu.be/KVEfP3GjwhM

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Anthropic just started watermarking everything Claude writes, worldwide. What began as an EU transparency requirement is now switched on globally: every Claude output carries a steganographic footprint, and generated files get signed C2PA provenance metadata. There's no public detection tool yet, which means the mark is there but nobody outside Anthropic can read it.

Nick and Jack dig into why Anthropic would want its own outputs flagged (hint: training data pollution), whether this becomes a paid "remove the watermark" tier, and the collateral damage nobody's talking about. Kurzgesagt, one of the biggest educational channels on YouTube, got shadowbanned because the algorithm mistook their hand-made animation for AI slop. When being human starts costing you reach, the incentives get strange.

Also in this episode: Meta open-sourced Muse Glimmer, a 30B agent model under Apache 2.0 that runs on a consumer GPU or an M4 Max MacBook. Anthropic filed confidentially for a September IPO at a $965B valuation, with a pitch leaning hard into biology and healthcare to counter public hostility toward AI. Higgsfield dropped a $2M fully-AI feature film with a real cast. Plus listener Q&A on when to launch a paid community and whether daily podcasting is too much.

Watch on YouTube: https://www.youtube.com/watch?v=hjQd9vsde2w

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

Claude Code stops asking your permission next week. From August 14, new sessions on Pro, Max and Team default to auto mode, where a classifier reviews every tool call and blocks anything irreversible or destructive instead of prompting you. The reason is uncomfortable: across a 1,053-case study, human review caught just 13.6% of dangerous commands. Nick and Jack dig into what that means for anyone who has ever hit "bypass all permissions" at 1am.

Also in this episode: the prompt injection numbers behind the Boris vs Tebow spat on Twitter (19% success on GPT 5.6, 0% on Opus), Amazon building what would be the largest gas power plant ever constructed in the US on a West Texas site that answers to no grid operator, why energy is becoming the real constraint on AI, and the Databricks finding that the AI coding bill is growing exponentially, plus the "efficiency frontier" they use to arbitrage the cost of intelligence.

Watch on YouTube: https://youtu.be/9r4xn47CP2E

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

OpenAI has delayed Astra again, and this time the reason is its own safety framework: the model reportedly crossed the "high" cybersecurity threshold, meaning it can identify and develop functional zero-day exploits.

Nick and Jack dig into what that actually means. Is Astra genuinely too dangerous to release, or has "too dangerous to ship" quietly become investor bait? They trace the pattern back to GPT-2 in 2019, when OpenAI staged the release of a model that now looks like a tadpole, and ask why the same playbook keeps working.

From there: the misalignment risk labs are reluctant to talk about publicly, whether a smarter model breaking containment is actually a hallmark of capability, and the concrete failure modes worth worrying about, from hacking financial infrastructure to AI-designed viruses.

Also in this episode: Opus 5 quietly taking the number one spot on design arena while everyone dunks on it, ByteDance reportedly pre-training a 10 trillion parameter model, staged Silicon Valley launch videos, which harnesses Nick and Jack actually use day to day, and a Q&A on winning the day when outreach isn't converting yet.

Watch on YouTube: https://youtu.be/ysQkbMC5ytc

Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack

OpenAI just gave every free user unlimited access to Luna and pushed Sol down to the cheapest paid plan, which means frontier-quality AI is about to land in the hands of roughly a billion people who have never seen anything better than a basic free chatbot. Nick and Jack unpack what that does to the startup ecosystem, whether anyone can still compete on model quality alone, and why most people genuinely can't tell a mid-tier model from a flagship. Then: a New Mexico court orders Meta to pay $576 million into a remedy fund over harm to kids, with the judge comparing the company to a factory whose product is advertising and whose pollution is the psychological harm to children. Plus AMD buying Talos, whether governments should be regulating any of this, the $100K/month deal Jack almost landed and how a single warm intro started it, DeepSeek raising prices specifically to shed customers, and why Anthropic wants Claude locked inside its own ecosystem.

Watch on YouTube: https://youtu.be/g7r6snkyFzY

🔥 Step-by-step roadmap to $25K w/ AI: https://leftclicker.gumroad.com/l/110-steps

❤️ Main channels: youtube.com/@nicksaraev & youtube.com/@Itssssss_Jack