Social Media Briefing — June 18, 2026

Following yesterday's News coverage, Gary Marcus argues LLMs are not reliable enough to be sound, says Trump's request for non-circumventable systems cannot be met, and frames it as a generative-AI-wide problem rather than an Anthropic one.

DrJim Fan gives a behind-the-scenes tour of Physical AutoResearch via the ENPIRE system, detailing two-layer hardware safety (kinematic limits and torque-limited compliant grippers) and tamper-proof reward definitions to stop agents gaming their goals during overnight unattended robot runs.

OpenAI introduces LifeSciBench, a benchmark co-developed with 173 scientists, comprising 750 expert-authored tasks across seven biological research workflows.

Gary Marcus criticizes Mark Zuckerberg, arguing AGI cannot be reached through data labeling and that turning Meta's research division into a data-labeling operation is a major corporate blunder.

OpenAI announces GPT-5.4 helped drive a medicinal chemistry project from literature review to a validated experimental result, partnering with a Maria AI system to propose an unexpected improvement to a widely used drug-discovery reaction.

Following yesterday's News coverage, Notes leaked financials suggest OpenAI has 40%+ gross margins on serving customers while training stays costly, and argues automating AI research could improve training efficiency.

Greg Brockman remarks that software engineering is so different now that it is hard to remember what it was like even six months ago.

Continuing our coverage from yesterday, TheRundownAI compiles news on Anthropic, the US government, and Mythos/Fable model access, including G7 talks, employee fears of being unfairly targeted, an expanding access list, and demands to fix jailbreak vulnerabilities before re-release.

Claude Code contributor argues we are entering an era where models generate correct code for more tasks, with the human role shifting to building guardrails, verifiers, and running model-plus-verifier loops on tasks.

Following yesterday's News coverage, Ethan Mollick notes that leaked financials suggest OpenAI is profitable on customer serving with 40%+ gross margins while training remains very costly, and argues automating AI research could boost training efficiency.

Burkov summarizes a recent Google arXiv paper showing transformers struggle with state tracking, using a higher-or-lower guessing game and word-sense ambiguity to illustrate failure to maintain and revise a running summary as context evolves.

Shares a creative AI benchmark asking models to build a procedurally generated 3D harbor-town evolution simulation, with a gallery comparing 20 models.