Sunday, May 24, 2026
Heterogeneous intelligence is the new scaling law.
May 24 · 4 videos
DeepMind gives employees worse quotas than customers.
Heterogeneous systems beat GPT 5.2 by 25 percent.
Every agent needs a human gardener.
Cloudflare is building the actor model into infrastructure.
The era of monolithic model dominance is ending.
“Automation is a lie. Every agent needs a human.”
How Google DeepMind Runs Agents at Scale — KP Sawhney & Ian Ballantyne, Google DeepMind
KP Sawhney · AI Engineer · 25 min
Watch on YouTube →KP Sawhney and Ian Ballantyne explain how Google DeepMind manages agentic workflows at scale. They prioritize external customer reliability over internal team usage.
- DeepMind uses a Darwinian skill library where only high performance tools are retained.
- Internal users face stricter token quotas than paying customers to ensure system stability.
- The Antigravity platform moves from monolithic context blobs to shared file systems.
- Agent Trajectory Stores provide observability for complex multi step reasoning.
- Mock TPUs allow testing agent logic without consuming expensive compute hours.
- Human monitoring teams intervene manually if internal usage spikes occur.
⚡️ Why you should build Science Fiction — Sunil Pai, Cloudflare
Sunil Pai · Latent Space · 14 min
Watch on YouTube →Sunil Pai discusses building stateful serverless infrastructure for AI agents at Cloudflare. He advocates for building ambitious projects instead of incremental enterprise tools.
- Durable Objects implement the actor model at the infrastructure layer rather than userland.
- Dynamic Workers allow zero latency execution of code generated by LLMs.
- The industry is searching for a standardized architecture to decouple intelligence from execution.
- Vendoring via forking is a pragmatic way to own dependencies and avoid breaking APIs.
- Maintainers are increasingly shutting down public contributions to avoid fake security reports.
- Cloudflare's API contains 2,600 endpoints for agents to navigate.
Scaling the Next Paradigm of Heterogeneous Intelligence — Adrian Bertagnoli, Callosum
Adrian Bertagnoli · AI Engineer · 15 min
Watch on YouTube →Adrian Bertagnoli presents the case for heterogeneous intelligence over monolithic models. He shows how specialized hardware and routing outperform frontier models.
- Diverse systems consistently outperform uniform ones under real world constraints.
- Callosum achieved an 18 to 25 percent performance lead over GPT 5.2 on benchmarks.
- Recursive language models on Cerebras hardware are 7x cheaper and 5x faster than frontier models.
- Routing subtasks to specialized chips like SambaNova reduces costs by 3.7x.
- The competitive advantage is shifting from training models to building orchestration layers.
- Vertical integration of intelligence and hardware represents the final stage of AI evolution.
The AI paradox: More automation, more humans, more work | Dan Shipper
Dan Shipper · Lenny's Podcast · 94 min
Watch on YouTube →Dan Shipper argues that AI automation creates more work by requiring humans to act as system gardeners. He predicts a shift toward Bring Your Own Token models in SaaS.
- The AI Paradox suggests that automated agents require human direction to remain valuable.
- GPT 5.5 scored 62 out of 100 on the proprietary Senior Engineer Benchmark.
- Current models can perform tasks for 17 hours autonomously with 50 percent accuracy.
- SaaS margins may improve as users bring their own tokens to applications.
- Human value is shifting from execution to framing and strategic direction.
- Every's headcount doubled to 30 people despite being an AI native company.
References
PeopleKP Sawhney · Ian Ballantyne · Sunil Pai · Adrian Bertagnoli · Dan Shipper · Kieran Klaassen (@kieranklaassen) · Brandon Gell · Marcus · Nitesh
ToolsAntigravity · Agent Trajectory Store · Durable Objects · Dynamic Workers · Cerebras · SambaNova · Qwen 3 · Gemini Ultra · Gemma · Claude Code · Codex