Marco Mascorro

@mascobot

Researching LLMs | Roboticist | Prev: Partner @a16z, cofounded Fellow Robots, @BMW Research

following
followers

Hardware and AI models as a fun combination

A user says hardware is fun, and AI models make it even better.

1 Source, 2d ago, first seen 2d ago

Language models' claimed ability to learn to detect when their activations are read

A user asks how models could learn to do this and points to a thread.

23 Sources, 4h ago, updated 4h ago

AI models may be more interpretable than previously thought

Goodfire researcher Nick Cammarata argues that modern tools reveal most neurons as legible in models up to trillions of parameters.

197 Sources, 1d ago, updated 1d ago

Third Moment Research launches with plans for enterprise AI and reinforcement-learning environments

A backer says the planned reinforcement-learning environments will start with applications in quantitative finance.

6 Sources, 6d ago, updated 6d ago

Conference submissions allegedly flooded with AI-generated papers

A post describes a video as allegedly showing someone gaming peer review by submitting AI-generated papers in volume.

16 Sources, 18h ago, updated 18h ago

Claims of high eight-figure to mid nine-figure run rates for data and RL environment companies in less than a year

A user says many investors’ surprise shows how little they understand about models, training and data.

5 Sources, 10d ago, updated 10d ago

TypeSafe AI's model launch comes with claims of trillions of tokens served daily

The poster says 29.4% of the Fortune 500 “showed up” and reports a Series A from a16z.

30 Sources, 10d ago, updated 10d ago

Karotte goes open source for building reinforcement-learning environments

Its developers say they used the framework for a year and hardened it through more than 1 million evaluation runs.

7 Sources, 12d ago, updated 12d ago

The case for scarce data’s rising value as AI models advance

A commenter says scarce data has “absurd and increasing value” as AI models reach narrow superintelligence.

6 Sources, 11d ago, updated 11d ago

Agent-first coding interfaces versus terminal tabs and IDEs

A user argues that juggling 30 terminal tabs adds cognitive overhead when coding-agent context needs to persist.

18 Sources, 13d ago, updated 13d ago

An interview with Anthropic technical leaders on the AI frontier

The interviewer says they asked hard questions about open source, regulatory capture and slowing the US relative to China.

10 Sources, 18d ago, updated 18d ago

Underdog introduces an on-device AI assistant and announces its backers

Underdog says its Mac assistant stores personal data on-device, with iOS, Windows and Android versions planned.

83 Sources, 10d ago, updated 10d ago

AMD agrees to acquire World Labs in $8.2 billion all-stock deal

World Labs says Fei-Fei Li will join AMD as executive vice president and chief scientist, working directly with CEO Lisa Su. The company says the move will bring its spatial AI research closer to the hardware.

93 Sources, 20d ago, updated 20d ago

An invitation to discuss robotics hardware, control and pretraining at IROS

A user welcomed IROS attendees to Pittsburgh and CMU and invited in-person chats with people working on hardware, arm or whole-body control, and ego/UMI pretraining.

3 Sources, 22d ago, updated 22d ago

The case for a ‘Millennium Prize’ for AI alignment

One user argues that better-aligned incentives could encourage independent research, calling MATS and the Ant Fellows Program good starting points.

1 Source, 29d ago, first seen 29d ago

Underdog creator says the AI operating system runs 100% locally

The creator says privacy fears drove the project, including a claim that anyone can get a friend's real Social Security number from Claude.

17 Sources, 38d ago, first seen 38d ago

The case for using just one AI coding agent

Pushing back on “software factories,” the user argues for letting a single AI coding agent spin up others if needed rather than building that setup yourself.