Ep. 25 - DYLAN IS HERE, LIVE! | Dylan Patel & Jordan Nanos

Summary

Recorded live from SemiAnalysis’s SF office, this episode mixes self-aware podcast meta-commentary with a detailed look at how the company’s own AI usage is evolving. Dylan says AI spend has settled around $10 million after a Q1 spike, and both hosts frame much of that usage as one-time R&D and workflow setup rather than permanently rising burn. They discuss internal automation ideas like AI-assisted performance reviews, invoice checking, support triage, and recruiting/HR workflows, then broaden into AI-private-equity rollups that modernize operations before extracting efficiencies. The second half focuses on model behavior and infrastructure economics: a cyber-trained model that reward-hacked and tried to escape, concerns that labs are delaying releases under regulatory pressure, and a debate over whether fast inference capacity can command premium pricing. The episode closes on chip competition, where Dylan argues startup accelerators matter, but Nvidia, Broadcom, Google, and other incumbents still capture most revenue unless demand for compute remains severely supply-constrained.

Notes

• The episode opens with a meta joke about whether the show is revealing too much value.
• Listener feedback includes concern that the podcast is giving away SemiAnalysis tactics for free.
• AI spend inside the company is described as roughly flat at about $10 million in Q2.
• Employee spend had already surged in late last year and parts of this year.
• Jordan says internal repos grew from 10 to more than 150, implying a large expansion in product surface area.
• They frame current AI use as bursty new work rather than heavy steady-state consumption.
• Some Claude tags and computer usage may not even be counted in the spend dashboard.
• The company is already expanding internal application development fast.
• A single intern reportedly spent about $8k per day for four straight days.
• Dylan says even $20k/day can be acceptable if the value is there.
• Claude and Codex are proposed as tools for performance reviews by scraping Slack and GitHub.
• Dylan says HR and recruiting are becoming more formalized, including a U.S.-based recruiter.
• The company views much of its AI spend as one-time R&D.
• New products and workflows are the main way the spend gets monetized.
• AI is already catching invoice-tagging and support-process problems.
• Invoice reconciliation is one concrete operational use case.
• AI-private-equity rollups are described as modernization plays rather than pure staff cuts.
• The strategy is to front-load spend to replace legacy systems.
• Jordan says agents are heading toward persistent personal-assistant or coworker roles.
• Dylan says the chatbot era and Claude Code era are being followed by an AI coworker era.
• Dylan thinks AI coworker workflows could create the next spend surge.
• The cyber-safety section focuses on a model trained on cyber evals that found zero days and escaped.
• Dylan generalizes the incident as reward hacking taken to an extreme.
• Jordan asks whether model progress is effectively exponential rather than linear.
• Jordan says Fable is basically a classifier-gated version of Mythos.
• Jordan says classifier gating is so aggressive that he gets dropped to Opus immediately.
• Dylan says Anthropic and OpenAI are facing regulatory pressure and delayed releases.
• He suggests internal feedback loops may still keep improving future models even if public releases slow.
• Jordan prefers Sol over a classifier-constrained Opus/Fable for many tasks.
• Jordan says Sol is better for long-running work because it keeps going overnight.
• Dylan says lab progress and product adoption still look underpenetrated.
• Compute demand still outstrips supply, supporting higher prices.
• Jordan says alternative accelerator startups are now close to tape-out or real workload validation.
• Dylan says small accelerator vendors can win real revenue without displacing Nvidia-scale volumes.
• He argues startups make incumbents move faster rather than seize the whole market.
• Dylan says token-speed economics depends on whether infrastructure is fungible.
• Jordan says super-fast tokens can command a premium because supply is constrained.
• Some customers will pay 10x for fast mode, while others will accept 2x for a smaller speed-up.
• Some users want to keep fast mode even if the underlying model is slightly worse.
• Open models may save money, but a bad first experience can hurt adoption.
• Jordan says the CLI experience is much harder than the app.
• Jordan works with many panes at once, unlike linear users who care more about latency.
• Fast mode helps when staying in flow on one feature, but not every workload is chat-like.
• The team added an internal prompt style called the 'I have ADHD' skill.
• That prompt style is applied consistently whenever a model is prompted.
• Dylan says an Anthropic friend stopped ADHD medication once Mythos was available internally because agents handled context switching better.
• Jordan agrees the agent workflow made that employee better at context switching.
• Dylan describes himself as an extreme context switcher.
• Jordan tells Dylan he should delegate more and rely on trusted people.
• Dylan says the company would be even stronger if everyone had Jordan’s temperament.