AI Twitter Digest: Mythos Reality Check, Big Tech’s Cash Crunch, and Shifting Bottlenecks — 2026-05-08#

Highlights#

Today’s AI discourse is caught between staggering capital expenditure and a sobering reality check on model capabilities. While Big Tech burns through cash to fund a projected $715 billion in 2026 AI infrastructure, the latest evaluations of Anthropic’s heavily-hyped Mythos model reveal an impressive but strictly on-trend tool rather than a quantum leap. Meanwhile, the strategic bottlenecks of software development are fundamentally shifting from coding to distribution as AI lowers the barrier to entry.

Top Stories#

• Big Tech’s AI Cash Burn: A new report projects that combined free cash flow for Microsoft, Alphabet, Amazon, Meta, and Oracle will drop over 70% to around $100 billion by the end of 2026 due to $715 billion in AI capital expenditures. Despite reporting record profits on paper, these companies are running low on actual cash, forcing them to project issuing an unprecedented $175 billion in new debt in 2026 alone. (Source)
• De-hyping Anthropic’s Mythos: Recent data from Mozilla and cybersecurity experts indicate that while Anthropic’s Mythos model is highly effective at finding bugs, it is not a “godlike” leap in capabilities. Experts note that similar software vulnerabilities can be found by existing OpenAI and Anthropic models, and the threat primarily targets poorly secured systems rather than critical infrastructure. (Source)
• OpenAI and a16z Super PAC Astroturfing: Investigations reveal sloppy astroturfing campaigns by the “Leading the Future” super PAC, which utilized fake AI-generated journalism platforms and obscure shell LLCs to hide disbursements. The campaign even paid TikTok creators up to $5,000 per video to promote pro-AI content without disclosing the financial backing. (Source)
• DeepMind Achieves Math Milestone: Google DeepMind introduced a multi-agent AI co-mathematician designed to actively collaborate with human experts on open-ended research mathematics. In autonomous mode evaluation, the system scored an unprecedented 48% on the rigorous FrontierMath Tier 4 problems, setting a new high score among evaluated AI systems. (Source)

Articles Worth Reading#

The New Bottlenecks of AI-Assisted Software (Source) Aaron Levie and Gergely Orosz note that as AI makes building software rapidly accessible, the fundamental bottlenecks for startups are shifting entirely to distribution. Because competing teams can now easily build similar AI products in parallel, the only true differentiators left are marketing, sales, and deep customer engagement. Furthermore, while AI helps start projects quickly, it creates a massive new workload for humans who must now figure out what the AI agents should do, review their output, and complete the tedious “last mile” of work that the models cannot finish.

The Dangers of AI “Fact-Checking” Offload (Source) Isaac Saul highlights a terrifying new trend where readers use ChatGPT to independently “fact-check” original reporting. In one instance, a reader copy-pasted a story about political corruption into the AI, which proceeded to confidently and falsely claim that documented, real-world events—such as the existence of the Trump family’s World Liberty Financial crypto firm—were entirely fabricated. This exposes a massive vulnerability in our information ecosystem as users increasingly offload critical thinking to LLMs that hallucinate when stripped of external web links.

HTML is the New Markdown (Source) Anthropic’s Thariq argues that with advanced coding tools, writing in markdown is becoming an obsolete practice. Instead of relying on limited formatting shorthand, developers can now efficiently use tools like Claude Code to generate rich, customized HTML directly for their workflows. This shift illustrates how generative coding tools are fundamentally reducing the friction of producing complex outputs, rendering intermediate formats completely unnecessary.