SnackOnAI Blog

Jun 2, 2026

Feynman: The AI Research Agent That Verifies Before It Summarizes

Every AI research tool today rushes to produce a summary. Feynman (companion-inc/feynman, MIT, 7k stars, April 2026) is built on the opposite philosophy: verify first, summarize second. It dispatches four specialized sub-agents in parallel (Researcher, Reviewer, Writer, Verifier), grounds every claim to a direct URL, and produces a structured research brief with live citation verification. The architecture is grounded in the source, not in the model's training data.

Jun 1, 2026

gstack: Why the Y Combinator CEO Turned His Claude Code Setup Into a Software Factory With 23 Specialist Roles

gstack (MIT, 105k stars, March 2026) is Garry Tan's published Claude Code configuration: 23 opinionated slash commands that assign specialist roles (CEO, Eng Manager, Designer, QA Lead, Security Officer, Release Manager, Doc Engineer) to Claude, cycling through a fixed Think → Plan → Build → Review → Test → Ship → Reflect loop. The design thesis is that Claude performs better with role identity and process structure than with free-form prompting, and the self-reported numbers are specific enough to be interesting: 600,000 lines of production code in 60 days.

May 31, 2026

Open-Generative-AI: The Free AI Studio That Is Not Actually Running Models on Your Machine

Open-Generative-AI (MIT, 17.5k stars, trending April 2026) is billed as a self-hosted, uncensored alternative to Higgsfield, Freepik, and Krea. It is genuinely useful. It is also an API aggregator with a polished Next.js frontend, not a local inference stack. Understanding exactly what runs where, what "free" means, and what the MuAPI dependency implies for production use is the analysis most coverage skips.

May 30, 2026

ADI Reasoning: The Symbolic Scaffold That Forces LLMs to Separate Hypothesis Generation From Verification

Chain-of-thought prompting lets LLMs perform abduction, deduction, and induction simultaneously in a single autoregressive pass, with no separation and no accountability for which mode is active at any step. The ADI Protocol formalizes Peirce's tripartite inference as an explicit scaffold, enforces consistency through five algebraic invariants (the Gamma Quintet), and uses the Weakest Link bound to ensure no conclusion can exceed the reliability of its least-supported premise.

May 29, 2026

JEPA: Why Predicting in Pixel Space Was the Wrong Goal All Along

Self-supervised learning has been dominated by two ideas: reconstruct masked pixels (MAE), or force representations of different views to be similar (DINO, BYOL, SimCLR). JEPA (Joint-Embedding Predictive Architecture) rejects both. It predicts abstract representations of masked regions, not pixels. This single architectural choice produces richer semantic features with 10x less compute than MAE and zero hand-crafted augmentations. Yann LeCun has been arguing for this design for decades. The empirical results are now here.

May 28, 2026

TurboQuant: The Quantization Algorithm That Actually Proves Its Distortion Rate Is Near-Optimal

Every quantization method claims minimal quality loss. TurboQuant (Google Research, ICLR 2026) is among the first to prove it: the distortion rate is within a constant factor of the information-theoretic lower bound. The proof comes with a two-stage algorithm that works online, requires zero per-vector quantization overhead, and directly addresses the KV cache memory bottleneck that limits long-context LLM inference.

May 27, 2026

MiniMax M2.7: The Model That Ran Its Own RL Experiments and Got 30% Better Without a Human Touching the Code

MiniMax M2.7 is not a model that was trained by engineers. It is a model that participated in training itself. An internal version of M2.7 ran over 100 autonomous rounds of scaffold optimization, evaluated its own outputs, decided which changes to keep, and achieved a 30% performance improvement on internal benchmarks. This is not a demo. It is the production pipeline that built the model you can use today.

May 26, 2026

RelBench v2: Four New Databases and What They Reveal About Where Relational Deep Learning Breaks Down

RelBench v2 does not just add more databases. It adds databases specifically chosen to stress-test relational deep learning in domains where the pkey-fkey graph hypothesis is hardest to satisfy: high-cardinality sparse interactions, long-tail distributions, and temporal dynamics that defeat simple neighborhood aggregation. The leaderboard results on the four new databases tell a more honest story than the headline benchmark numbers.

May 25, 2026

RelBench v1: The Benchmark That Forced Honest Evaluation on Relational Deep Learning

Every published result on relational database ML before RelBench was incomparable: different temporal splits, different leakage handling, different metrics. RelBench v1 fixed all three simultaneously by making correct temporal evaluation the default behavior, not the careful choice. The benchmark is the infrastructure. The databases are the test suite. The enforced defaults are the contribution.

May 24, 2026

FST: The Dual-Engine Training Method That Reaches Peak Performance With Three Times Fewer Steps

Reinforcement learning trains LLMs by updating parameters. Prompt optimization adapts LLMs by updating context. Everyone picks one. Fast-Slow Training (FST) runs both simultaneously, treating the prompt as fast weights that absorb task-specific information and the parameters as slow weights that preserve general reasoning, reaching higher performance in fewer steps while maintaining the model's ability to keep learning.

May 23, 2026

R2Code: Why Your LLM Knows What Code to Write But Not Which Requirement It Satisfies

LLM-generated code has a traceability problem. The model produces code that works, but cannot reliably tell you which requirement each function implements, which requirement has no code at all, and which code has no requirement to justify it. R2Code is the self-reflective framework that closes this gap with an iterative generate-verify-reflect loop and outperforms prior approaches on precision, recall, and F1 across standard benchmark datasets.

May 22, 2026

Smolagents: The Agent Framework That Proves JSON Tool Calling Was the Wrong Abstraction All Along

Every major AI framework ships agents that describe tool calls as JSON objects. Smolagents ships agents that write Python. This is not a superficial difference. Python is a better language for expressing actions than JSON, and the research agrees. Smolagents is the framework that takes this seriously, keeps the entire implementation under ~1,000 lines, and benchmarks the result.

May 19, 2026

Kunlun: Why Meta's Ads Models Are Wasting 83% of Their GPU, and How They Fixed It

Recommendation system models at Meta achieve 3-15% Model FLOPs Utilization on the same NVIDIA B200 GPUs where LLMs achieve 40-60%. This is not a scaling problem. It is an efficiency problem. Kunlun is the architecture that fixes it, raising MFU from 17% to 37%, doubling scaling efficiency, and establishing predictable power-law scaling for one of the most economically important ML workloads on the planet.

May 18, 2026

RepForge: The Tool That Watches Claude Code Build Your App and Quizzes You on the CS You Missed

Every developer using AI coding agents has the same silent problem: the code ships, the PR merges, and the developer has no idea why the implementation works. RepForge is the tool that sits beside Claude Code and Codex, extracts the computer science concepts from each session, and turns them into spaced repetition review challenges before you forget them. The learning happens. You just have to show up for the review.

May 17, 2026

BountyBench: The First Cybersecurity Benchmark That Measures Dollar Impact, Not Just Success Rate

Bug bounties pay real money for real vulnerabilities. BountyBench is the first cybersecurity AI benchmark that inherits this economic framing: every task has a dollar value attached, every success rate maps to a bounty total, and the headline result is not "67.5% exploit rate" but "$14,422 in defended patches." The unit of measurement is the correct one.

May 16, 2026

CyberGym: The Benchmark Where AI Agents Try to Break Real Software, and Mostly Fail

The best AI agent on CyberGym, a benchmark of 1,507 real-world vulnerabilities from production software, achieves a 22% success rate. That number has two implications: AI agents are already capable enough to reproduce one in five real vulnerabilities autonomously, and four in five vulnerabilities remain beyond their reach. CyberGym is the first benchmark large enough and realistic enough to make both implications defensible.

May 15, 2026

LTX-2: The First Open-Weights Model That Generates Video and Audio in One Pass

Every text-to-video model released before LTX-2 generates silent video. The audio you see in demos is added afterward by a separate model or manually. LTX-2 (Lightricks, arXiv:2601.03233, January 6, 2026) generates synchronized video and audio jointly in a single diffusion pass. The architecture required to make that work, an asymmetric dual-stream transformer with 14B video parameters and 5B audio parameters, is the story.

May 14, 2026

NNGPT: The AutoML System That Writes, Runs, Judges, and Improves Its Own Neural Networks

The traditional AutoML search loop wastes thousands of GPU hours evaluating candidate architectures one by one. NNGPT replaces most of that search with a single LLM prompt, then closes the loop: every generated network that runs gets fed back to improve the model that generated it. The system has already produced over 10,000 validated architectures.

May 13, 2026

Axolotl: One YAML File to Rule Every Fine-Tuning Method That Exists

You do not need to understand DeepSpeed internals, Flash Attention dispatch logic, or FSDP shard mechanics to train a competitive fine-tuned model. Axolotl wraps all of it in a single YAML config. The discipline required to make that abstraction not leak is the entire engineering story.

May 12, 2026

Hermes Agent: The OpenClaw Fork That Fixed Every Problem OpenClaw Left Unsolved

Hermes Agent is what OpenClaw becomes when the team building it also trains the models powering it. The architectural differences are not cosmetic. They are the result of NousResearch shipping production agent infrastructure and then open-sourcing the correct version.

May 11, 2026

OpenClaw: The 371k-Star Agent Framework That Proves the Architecture Was Obvious All Along

OpenClaw is not a new idea. It is an obvious idea that no one packaged correctly until a weekend project in November 2025. The architectural pattern it implements, persistent workspace files as agent memory, a heartbeat cron loop for proactivity, and a skills registry for extensibility, is now the reference architecture for personal AI agents. That is worth understanding precisely.

May 10, 2026

KARL: Why Databricks Built a Custom RL Agent Instead of Paying for Claude

Enterprise search agents that route every query through frontier model APIs are not a sustainable architecture. KARL is the result of Databricks asking what happens when you train a purpose-built model for exactly this problem, and the out-of-distribution generalization results are the most important numbers in the paper.

May 9, 2026

Autoresearch: The Engineering Behind Karpathy's Autonomous ML Experiment Loop

Autonomous ML experimentation is not a capability problem waiting to be solved. It is a systems design problem that autoresearch has already solved with three files, a fixed time budget, and an immutable judge. The human programs the research direction. The agent runs the science.

May 8, 2026

AGENTS.md: The README for AI Agents That Usually Makes Things Worse

The file format adopted by 60,000 repositories to guide AI coding agents actually reduces performance when generated by an LLM, increases inference costs by 20-23%, and leads to more thorough but less effective agent behavior. The research is unambiguous. Most teams are using AGENTS.md wrong.

May 7, 2026

TinyAGI: The Agent Teams Orchestrator Built for the One-Person Company

Most multi-agent frameworks are built for research labs or enterprise teams. TinyAGI is built for one person who wants to run a company with AI. The architectural decisions that follow from that goal are different in every layer of the stack.

Mohinish S

Serverless Ventures | Cloud, Data & Distributed Systems | Angel & Advisor | Infra & Data Startups