The Best Way to Talk to Your AI Agents

The episode opens with headlines about Anthropic reportedly considering a fundraising round that would value the company at $900 billion pre-money — surpassing OpenAI's $852 billion valuation — with sources saying demand is overwhelming and compute concerns have been eased by a SpaceX deal. A pre-IPO blockchain instrument was trading at an implied $1.2 trillion valuation, suggesting speculative appetite is even higher.

The host then covers Cerebrus's upcoming IPO, where demand has surged to 20x available shares, prompting the company to consider raising its price range from $115–$125 to $150–$160 per share, implying a valuation jump from $26B to over $34B. Analysts are mixed, with some bullish on hype and others warning of fundamental risks.

On the chip side, TSMC reported its slowest sales growth in six months — 17.5% annualized, roughly half of analyst forecasts — attributed to a slowdown in non-AI consumer electronics and physical capacity constraints on advanced AI chip manufacturing. Meanwhile, Apple signed a preliminary chip-making agreement with Intel, partly at the White House's urging, diversifying away from TSMC. AMD and Intel both surged ~25% in a week, while Nvidia lagged at 8%. A speculative note covered household micro data centers being explored by housing developers in partnership with Nvidia as a distributed compute solution.

The main episode focuses on a viral post by Tariq Shahipar, an Anthropic engineer on the Claude Code team, arguing that HTML should replace Markdown as the dominant format for AI agent communication. Tariq's case for HTML rests on information density (tables, CSS, SVG), visual clarity (tabs, diagrams, mobile responsiveness), ease of sharing, two-way interactivity (sliders, knobs), and subjective enjoyment. The post reached ~10 million views and sparked significant debate.

Critics argued HTML uses far more tokens than Markdown, raising costs — with some cynics suggesting Tariq, as an Anthropic employee, was incentivized to promote token-heavy formats. More measured responses noted that the choice depends on audience (humans vs. agents), lifecycle (edited many times vs. written once), and document longevity. The host synthesizes these views but goes further, arguing the real significance is a fundamental shift in the nature of knowledge work.

The host's central thesis is that in the agentic era, the 'atomic unit of knowledge work' has shifted from producing a final output to 'staging conditions' for an agent to produce that output. He describes spending significant time in a 'liminal space' — handing off context between sessions, tools, and agents — where documents exist in a state of 'mixed doneness': some parts locked, some open, some provisional. Markdown forces excessive meta-commentary to encode these states, whereas HTML's native features (color-coding, tabs, expandable sections, visual hierarchy) can communicate which parts are firm, exploratory, or undecided without cluttering the document itself. The host frames this as an early-stage exploration of how to work in a new modality where calibrating how much structure to impose on an agent — not over-specifying or under-specifying — is the critical new skill.