Big Tech’s $3T AI Bet, Claude Writes 90% of Anthropic’s Code, AIs Editing AI Brains | EP #308
Use AI to tackle a specific, measurable goal: define a benchmark, test performance, and improve the workflow over time.
- Create a benchmark for a problem you care about so AI progress can be measured and directed.
- For organizations over 50 people, move workflows to a separate AI stack one at a time and track how quickly each improves.
- Treat connected legacy systems as security risks; review internet-exposed industrial controls and update or isolate them.
The episode is a broad discussion; the notes capture its main claims and practical advice. · Best for: AI-curious professionals, business leaders, investors, and technology policy watchers
A wide-ranging discussion of AI infrastructure spending, job and everyday-task benchmarks, AI-assisted software development, direct neural-network weight editing, cybersecurity, space, health, and social change. The most actionable themes are measuring AI performance with task-specific benchmarks, migrating workflows incrementally, and addressing security gaps in connected systems.
AI, elections, and consumer devices0:00
- The hosts discuss using AI agents to research voting choices and raise concerns about their influence on elections.
- They debate whether phones and laptops may become thin clients that rely on remote compute, if connectivity becomes reliable enough.
- Apple’s reported device plans and memory costs prompt discussion about its AI strategy.
AI infrastructure spending and financing15:46
- The episode says five major technology companies have committed more than $3 trillion to AI infrastructure, including buildings, power, equipment, and leases.
- The panel debates whether financing and insurance capital can support further data-center investment, while noting risks around debt and uncertain returns.
- Participants expect continued demand for compute but acknowledge that new computing methods could change the economics.
Benchmarks for expert work and everyday tasks29:23
- GDPX is presented as a benchmark of long, realistic tasks across fields such as finance, law, biology, and software engineering; the models discussed score relatively low.
- LifeVal tests everyday assistance tasks such as budgeting, finding childcare, and dealing with government agencies.
- The panel argues that benchmarks help measure progress and suggests creating one for a domain or goal you care about.
AI-written code and recursive self-improvement49:10
- The episode reports that Claude wrote more than 90% of Anthropic’s new code in September 2026, and that internal Claude use grew sharply.
- The panel discusses AI systems improving models and company workflows, and recommends moving workflows over incrementally rather than replacing everything at once.
- Participants suggest tracking how quickly a system improves its own capabilities and processes.
Models editing neural-network weights1:01:58
- The hosts discuss MIT CSAIL research in which a frontier model diagnosed and edited a small open model’s weights, improving its score on a math task without retraining.
- They describe possible uses including adding capabilities to models and studying their internal structure, while noting the work is early and could also enable removal of safeguards.
- The group speculates that specialized model components could eventually be shared and attached to existing models.
Cybersecurity, science, and space1:13:53
- A cyberattack on remotely managed supermarket refrigeration is used to illustrate risks in internet-connected industrial systems.
- The episode covers a reported month-long containment of a positron and NASA’s proposed Artemis mission and lunar-base plans.
- The panel also discusses why AI could help modernize older industrial control systems.
Health, work, and audience questions1:34:58
- The health segment highlights sleep, diet, exercise, and mindset as practical foundations for brain health.
- In response to audience questions, speakers discuss finding purpose beyond employment, trade schools, data centers, empathy, and robotics.
- The closing discussion suggests AI could help people pursue personal goals and improve daily planning.
Step by step
- 1. Choose a concrete problem or workflow you want AI to improve.
- 2. Define a benchmark with clear tasks and a way to score results.
- 3. Test current AI tools against the benchmark and record cost, time, and task success.
- 4. For a larger organization, set up an AI workflow alongside existing operations and migrate tasks one at a time.
- 5. Re-run the benchmark regularly and track the rate of improvement.
- 6. Review internet-connected industrial controls and reduce unnecessary exposure; the episode specifically describes taking exposed systems off the open internet.
Tools & resources mentioned
- GDPX — benchmark discussed for extended, high-value professional tasks.
- LifeVal — OpenAI benchmark discussed for everyday assistant tasks.
- Claude — Anthropic coding model discussed in connection with internal code contributions.
- Gemini and GPT — models compared in the benchmark discussions.
- MIT CSAIL weight-editing research — study discussed on models directly editing another model’s weights.
- Hugging Face — model-sharing platform used as an analogy for a future marketplace for neural-network components.
- Grok and eBay — example tools in a proposed workflow for pricing and listing household items.
Notes are AI-generated from the video's transcript and may miss nuance.