Pretrained

Pierce and Richard are back for the second listener mailbag. They break down what reward hacking really is and why models so often learn the wrong lesson, explain practical fine-tuning (from pre-training to prompting), unpack why LLMs use tokens instead of words, how context length is a hardware versus mathematic limitation, and much more.

Pierce and Richard are joined by Bella Cooper-Brown, Oxford material scientist turned science fiction author. They cover why defense tech companies name themselves after Lord of the Rings, Asimov's three laws of robotics, why engineers don't read enough fiction, the intersection of material science and AI research, how speculative fiction shapes the tech we build, and the ethics of AI companions and relationships.

📚 Frankenstein; or, The Modern Prometheus by Mary Shelley (1818)
📚 Snow Crash by Neal Stephenson (1992)
📚 I, Robot by Isaac Asimov (1950)
📚 Klara and the Sun by Kazuo Ishiguro (2021)

Pierce and Richard cover the news that dropped over the holiday break. Getting breaking news incorporated within chatbots, OpenAI's "code red" over Google's Gemini 3, benchmarking the reliability of chain of thought to introspect model behavior, and a review of Claude Skills.

Further reading:
- https://www.wired.com/story/us-invaded-venezuela-and-captured-nicolas-maduro-chatgpt-disagrees
- https://fortune.com/2025/12/02/sam-altman-declares-code-red-google-gemini-ceo-sundar-pichai/
- https://openai.com/index/evaluating-chain-of-thought-monitorability/
- https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview

We should really be on Polymarket. Pierce and Richard make their bets on GPT-6, competition between the different letters of FAANG, dynamic websites calibrated to user preferences, and increasing quality of OSS models.

It's been a long year in the world of AI. Benchmarks are now almost totally saturated; the financial bubble keeps growing; spending more on inference compute; increasing competition from open source models; agents finally reach the mainstream; the frankly horrible job market for people out of school; multi-model models are back and increasing converging on transformer architectures. We cover them all in our anything goes holiday recap show. Plus - all the things that didn't happen this year.

Pierce reflects on his own 2025. Thoughts on choosing the right buyer persona, scaling an AI business from zero lines in a github repo, the feeling of finally reaching product market fit, boring versus interesting businesses, and more.

Richard takes the hot seat for the first episode of our 2025 recap series where we spend the rest of December looking back on what this year meant to us personally and in the world of AI/ML. We cover what it's like to defend a thesis in the UK, the difficulty of training meta-learning models, choosing a well scoped research topic, how to define small models, and what's needed to make them better.

OpenAI is rolling out shopping support to their users and plotting an ads rollout to challenge Google's ad business, we get a peek behind the curtain on SOTA image generation models with the release of Alibaba's Z-Image (and speculate this might be how nano banana has great text performance), and OpenReview exposes the identities behind double blind reviews.

Pierce and Richard cover OpenAI's new long range model compression in Codex, initial takeaways of Gemini 3.0 and Nano Banana Pro, Nvidia chip exports to the UAE and Saudi Arabia, and Cloudstrike's global outage. Plus - why Pierce prefers chicken to turkey.

Richard and Pierce break down the 5 levels of autonomy, whether Elon has a point about RGB vs lidar systems, sensor fusion algorithms, end to end learning in driving simulations, and more.