Tom differentiates between AI engineers who build models from research versus agent engineers who integrate models into applications.

Regulated industries and government customers adopt on-prem inference for compliance, opting out of cloud solutions to meet data residency and privacy...Show more

Despite being less capable, local models are increasingly viable as handheld hardware approaches supercomputer performance, enabling offline AI use ca...Show more

Their research paper emphasizes improving predictability and traceability to make LLM outputs more deterministic and auditable.

Tom argues Meta is quietly moving away from open-sourcing large language models after partners like Alibaba advanced Llama-based variants, instead foc...Show more

Claude offers more consistent and effective reasoning than faster or alternative LLMs, making it preferable for complex problem-solving despite speed ...Show more

Faster, simpler model responses can outperform deeper, research-heavy outputs for straightforward tasks, as exhaustive analysis may stall basic interp...Show more

Individual feedback often doesn't impact AI providers' service offerings, as they prioritize broader market factors over single-user opinions.

AI inference providers often trade off intelligence quality for faster speeds, with some providers showing up to 10% variation on benchmarks despite u...Show more

As AI agents mature, we’ll see web architectures evolve to natively support agent interactions, removing hacky workarounds like X/Y coordinates and sc...Show more

Weigh using broad, general-purpose agent tools like “computer use” against more specialized AI tools to balance flexibility with task-specific efficie...Show more

Tom Spencer argues that driving the incremental cost of automated web interactions to near zero will break current paradigms and unlock entirely new b...Show more

The speakers emphasize that tools like Playwright and other browser automation frameworks are highly complex and customizable, enabling developers to ...Show more

Browser-based AI agents can be easily exploited via prompt injection on websites, creating security and identity risks that concern infosec teams.

Simple system-level safety prompts can be prompt-injected or hacked, so relying solely on them may not prevent unwanted agent behaviors.

AI agent benchmarks using trivial tasks are "cutesy" and don't reflect real enterprise applications with high stakes and sensitive client data.

Daily computer tasks are dominated by manual form and CRUD interfaces, presenting substantial automation opportunities.

We need structured tooling and demos to ‘corral this new way of working’ in AI development instead of ad-hoc experimentation.

Tom Spencer observes that some AI pair-programming agents will ignore explicit code comments and enforce their own defaults, highlighting a trust issu...Show more

Cameron warns that widespread discussion of AI trends doesn't guarantee their validity and recommends thorough research before adoption.