Ion Stoica

@istoica05

Professor at UC Berkeley, co-founder of Databricks, Anyscale, Arena, Conviva.

following
followers

Exo explores AI models optimizing tasks by rebuilding their own harnesses

Exo's creators say models can use runtime logs and modify and rebuild their harness while optimizing a task.

4 Sources, 3d ago, updated 3d ago

JITEngine announced for model-specific LLM inference on Apple Silicon Macs

Promoters describe SkyDiscover-built engines tailored to a Mac and model, and announce six pre-built versions.

4 Sources, 5d ago, first seen 5d ago

Evolving Agent Teams reorganizes coding agents when they get stuck

Evolving Agent Teams' supervisor diagnoses dead ends and draws on saved lessons and team history to restructure coding agents.

6 Sources, 7d ago, updated 7d ago

Mirendil introduces Lor, an AI research product in private beta

Mirendil says Lor's agents write its code, run experiments and iterate on results for days.

15 Sources, 11d ago, updated 11d ago

FrontierCS 2 launches 48 paper-based computer science tasks for AI agents

Wenhao Chai says the team built its tasks from published papers and refined them with field experts.

16 Sources, 17d ago, updated 17d ago

Warp Sciences launches with a $215 million Series A for AI and robotics in medicine development

Co-founder Kevin Weil says robotic labs will generate data to train AI models for science.

123 Sources, 13d ago, updated 13d ago

Letta introduces FrontierMemory to score how memory from past agent attempts helps solve later tasks

Letta publishes separate model and agent-harness results for a test built around later task completion.

14 Sources, 19d ago, updated 19d ago

Researchers introduce TokenGate to prune tokens during long-context prefill

The authors report faster time-to-first-token and lower peak memory across diverse tasks, without accuracy degradation.

15 Sources, 18d ago, updated 18d ago

Mirendil unveils first pretrained model, claims it took one researcher 3.5 months

In the announcement, @bneyshabur places Mirendil-Base in the 100B–500B-parameter range and promises independent evaluations.

74 Sources, 14d ago, updated 14d ago

Readable Intelligence launches

A post congratulates @carrieeeeet and the team on the launch.

10 Sources, 21d ago, updated 21d ago

DuoDiscover is designed to find code optimization targets and cheap proxy evaluators together

A Google team member claims the agent ranked #1 on NVIDIA's SOL-ExecBench and reached 2× human experts on SWE-fficiency.

12 Sources, 24d ago, updated 24d ago

Student releases TPU kernels for modern linear attention variants

The student's instructor calls the kernels “impressively fast” and asks Google DeepMind about collaborating on model architecture research.

6 Sources, 25d ago, updated 25d ago

Arena launches AI agent Alignment Index and raises $200M at a $3.1B valuation

Arena says its index tracks unauthorized actions, false attribution and deceptive completion across 90,000-plus agent sessions and 27 models.

30 Sources, 30d ago, updated 30d ago

Arena's Series B: $200M at $3.1B, and a New Way to Measure AI Alignment

YouTube

Ion Stoica

Exo explores AI models optimizing tasks by rebuilding their own harnesses

JITEngine announced for model-specific LLM inference on Apple Silicon Macs

Evolving Agent Teams reorganizes coding agents when they get stuck

Mirendil introduces Lor, an AI research product in private beta

FrontierCS 2 launches 48 paper-based computer science tasks for AI agents

Warp Sciences launches with a $215 million Series A for AI and robotics in medicine development

Letta introduces FrontierMemory to score how memory from past agent attempts helps solve later tasks

Researchers introduce TokenGate to prune tokens during long-context prefill

Mirendil unveils first pretrained model, claims it took one researcher 3.5 months

Readable Intelligence launches

DuoDiscover is designed to find code optimization targets and cheap proxy evaluators together

Student releases TPU kernels for modern linear attention variants

Arena launches AI agent Alignment Index and raises $200M at a $3.1B valuation

High-quality video generation is claimed to run on hardware you already own

Does a Claude model really need Claude Code?

SkySynth's creators announce a plugin for Claude Code and Codex

FreeToken Runs Frontier Models on Consumer GPUs

Ion Stoica

Exo explores AI models optimizing tasks by rebuilding their own harnesses

JITEngine announced for model-specific LLM inference on Apple Silicon Macs

Evolving Agent Teams reorganizes coding agents when they get stuck

Mirendil introduces Lor, an AI research product in private beta

FrontierCS 2 launches 48 paper-based computer science tasks for AI agents

Warp Sciences launches with a $215 million Series A for AI and robotics in medicine development

Letta introduces FrontierMemory to score how memory from past agent attempts helps solve later tasks

Researchers introduce TokenGate to prune tokens during long-context prefill

Mirendil unveils first pretrained model, claims it took one researcher 3.5 months

Readable Intelligence launches

DuoDiscover is designed to find code optimization targets and cheap proxy evaluators together

Student releases TPU kernels for modern linear attention variants

Arena launches AI agent Alignment Index and raises $200M at a $3.1B valuation

High-quality video generation is claimed to run on hardware you already own

Does a Claude model really need Claude Code?

SkySynth's creators announce a plugin for Claude Code and Codex

FreeToken Runs Frontier Models on Consumer GPUs

Featured in

Exo explores AI models optimizing tasks by rebuilding their own harnesses

JITEngine announced for model-specific LLM inference on Apple Silicon Macs

Evolving Agent Teams reorganizes coding agents when they get stuck

Mirendil introduces Lor, an AI research product in private beta

FrontierCS 2 launches 48 paper-based computer science tasks for AI agents

Warp Sciences launches with a $215 million Series A for AI and robotics in medicine development

Letta introduces FrontierMemory to score how memory from past agent attempts helps solve later tasks

Researchers introduce TokenGate to prune tokens during long-context prefill

Mirendil unveils first pretrained model, claims it took one researcher 3.5 months

Readable Intelligence launches

DuoDiscover is designed to find code optimization targets and cheap proxy evaluators together

Student releases TPU kernels for modern linear attention variants

Arena launches AI agent Alignment Index and raises $200M at a $3.1B valuation

High-quality video generation is claimed to run on hardware you already own

Does a Claude model really need Claude Code?

SkySynth's creators announce a plugin for Claude Code and Codex

FreeToken Runs Frontier Models on Consumer GPUs

Featured in

Exo explores AI models optimizing tasks by rebuilding their own harnesses

JITEngine announced for model-specific LLM inference on Apple Silicon Macs

Evolving Agent Teams reorganizes coding agents when they get stuck

Mirendil introduces Lor, an AI research product in private beta

FrontierCS 2 launches 48 paper-based computer science tasks for AI agents

Warp Sciences launches with a $215 million Series A for AI and robotics in medicine development

Letta introduces FrontierMemory to score how memory from past agent attempts helps solve later tasks

Researchers introduce TokenGate to prune tokens during long-context prefill

Mirendil unveils first pretrained model, claims it took one researcher 3.5 months

Readable Intelligence launches

DuoDiscover is designed to find code optimization targets and cheap proxy evaluators together

Student releases TPU kernels for modern linear attention variants

Arena launches AI agent Alignment Index and raises $200M at a $3.1B valuation

High-quality video generation is claimed to run on hardware you already own

Does a Claude model really need Claude Code?

SkySynth's creators announce a plugin for Claude Code and Codex

FreeToken Runs Frontier Models on Consumer GPUs