Mohinish S

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

SnackOnAI Blog

The Postgres Predicate That Reads Your Table Twice

Oct 5, 2026

•

14 min read

The Postgres Predicate That Reads Your Table Twice

A WHERE clause that calls a model once per row would be useless. pg-jev makes it fast by secretly scanning your table ahead of the executor, and that read-ahead window, not your LIMIT, is what you actually pay for.

Mohinish S

SnackOnAI Blog

Jev Gives Up Writing And Gets A Type System

Sep 30, 2026

•

15 min read

Jev Gives Up Writing And Gets A Type System

A model that returns no text cannot hallucinate a citation. It can still hand your code a confidently wrong enum, and the confidence score it ships with is a formula you can write in one line.

Mohinish S

SnackOnAI Blog

MetaGPT Wrote The Org Chart Into The Prompt, Then Quietly Retired Half Of It

Sep 19, 2026

•

14 min read

MetaGPT Wrote The Org Chart Into The Prompt, Then Quietly Retired Half Of It

The paper that taught agents to pass documents instead of chatting is still right about the documents. Its own repository no longer believes in the assembly line.

Mohinish S

SnackOnAI Blog

BMAD Method Compiles Its Prompts, In Five Skills Of Thirty

Sep 18, 2026

•

13 min read

BMAD Method Compiles Its Prompts, In Five Skills Of Thirty

The flagship skill loads four hundred tokens to reach nineteen thousand, because a Python renderer builds the workflow on demand. Twenty five of the thirty skills still inline everything.

Mohinish S

SnackOnAI Blog

Orca Supports Twenty Nine Agents With One Terminal

Sep 16, 2026

•

13 min read

Orca Supports Twenty Nine Agents With One Terminal

Twenty nine agents needed a terminal. Windows process spawning needed six dedicated documents. The orchestration was never the hard part.

Mohinish S

SnackOnAI Blog

Sep 15, 2026

•

13 min read

DeepSeek Harness Ships Hot Reload In One Profile Of Five

Everything is a plugin, including the agent loop. Plugin discovery is a GitHub topic search. One runtime invariant is the only thing standing between those two facts.

Mohinish S

SnackOnAI Blog

Obscura Runs Every Page Through One JavaScript Thread

Sep 12, 2026

•

13 min read

Obscura Runs Every Page Through One JavaScript Thread

Thirty megabytes against two hundred is not a smaller Chrome. It is a different browser that implements less, serializes more, and needed six safety mechanisms to survive the choice.

Mohinish S

SnackOnAI Blog

llmfit's Fallback Estimate Can Be Fourteen Times Too Slow

Sep 11, 2026

•

13 min read

llmfit's Fallback Estimate Can Be Fourteen Times Too Slow

Decode is memory bound and prefill is compute bound, so Q4 halves one and leaves the other exactly where it was. Both numbers come from formulas the project publishes, which is what makes them checkable.

Mohinish S

SnackOnAI Blog

OSWorld's Hardest Baseline To Beat Is An Agent That Gives Up

Sep 3, 2026

•

14 min read

OSWorld's Hardest Baseline To Beat Is An Agent That Gives Up

Thirty of the three hundred sixty nine tasks are impossible. An agent that answers "cannot be done" to every single one scores higher than twenty one of the twenty six baselines in the paper.

Mohinish S

SnackOnAI Blog

Model Optimizer's Own Benchmark Says Four Bit Is Slower

Sep 2, 2026

•

13 min read

Model Optimizer's Own Benchmark Says Four Bit Is Slower

INT4 AWQ runs Llama3.1-8B seventeen percent slower than the unquantized baseline at batch sixty four, and the number is in NVIDIA's repo.

Mohinish S

SnackOnAI Blog

CUA-Lite's Parallelism Number Is Just The Memory Ratio

Sep 1, 2026

•

14 min read

CUA-Lite's Parallelism Number Is Just The Memory Ratio

Frontier agents passed OSWorld's human baseline sometime in the last year. CUA-Lite is not built to measure that. It is built to make the training loop cheap.

Mohinish S

SnackOnAI Blog

OpenSEO Is Very Good At Three Percent Of AI Citations

Aug 30, 2026

•

14 min read

OpenSEO Is Very Good At Three Percent Of AI Citations

AI engines cite peer and competitor pages twenty six times more often than they cite your own domain. Every SEO tool you own works on the domain.

Mohinish S

SnackOnAI Blog

Anthropic's Alignment Agents Did Not Out-Think Anyone

Aug 29, 2026

•

15 min read

Anthropic's Alignment Agents Did Not Out-Think Anyone

A full automated research run costs about what one human idea costs, and produces a hundred and seventy five of them. That is the whole result.

Mohinish S

SnackOnAI Blog

Aug 26, 2026

•

13 min read

Qwen3-VL And The Multimodal Tax

An eight billion parameter model with tools beat the flagship without them. The report has the numbers and does not connect them.

Mohinish S

SnackOnAI Blog

Rex-Omni, Or What Reinforcement Learning Actually Fixed

Aug 25, 2026

•

13 min read

Rex-Omni, Or What Reinforcement Learning Actually Fixed

A day of GRPO on eight GPUs delivered the headline gain. It improved tight box precision by one tenth of a point.

Mohinish S

SnackOnAI Blog

FreeToken, Or Why Your Laptop Was Never The Bottleneck

Aug 24, 2026

•

15 min read

FreeToken, Or Why Your Laptop Was Never The Bottleneck

An eight gigabyte laptop GPU just decoded a 35B MoE model at ninety two percent of an RTX 4090's rate, and the reason has nothing to do with the GPU.

Mohinish S

SnackOnAI Blog

LinkedIn Benchmarked Nine Memory Systems For Its Hiring Agent, Seven Lost To Pasting Everything Into The Prompt

Aug 11, 2026

•

15 min read

LinkedIn Benchmarked Nine Memory Systems For Its Hiring Agent, Seven Lost To Pasting Everything Into The Prompt

LinkedIn published a table this spring that quietly indicts most of the agent memory market. Their own paper, their own benchmark, their own baselines. Almost nobody has read the rows that matter.

Mohinish S

SnackOnAI Blog

Autograd Is Slow Per Operation And Free Per FLOP, And That One Sentence Tells You When To Use It

Aug 9, 2026

•

13 min read

Autograd Is Slow Per Operation And Free Per FLOP, And That One Sentence Tells You When To Use It

The reverse-mode engine that taught the field how to build autodiff is under two hundred lines of Python. You can read all of it in an afternoon. What you cannot do is make it fast, and the reason is architectural, not a missing optimization.

Mohinish S

SnackOnAI Blog

The Seedance Paper Is a Scoreboard, Not a Blueprint

Aug 8, 2026

•

14 min read

The Seedance Paper Is a Scoreboard, Not a Blueprint

ByteDance shipped a twenty six page paper on Seedance 2.0 containing twenty three pages of benchmark tables and zero equations.

Mohinish S

SnackOnAI Blog

AnyDoc Is Not Faster Than Docling. It Is Solving A Different Problem, And Almost Nobody Noticed

Aug 7, 2026

•

12 min read

AnyDoc Is Not Faster Than Docling. It Is Solving A Different Problem, And Almost Nobody Noticed

anydoc beats Docling by a factor of one hundred on document conversion speed. That number is real, reproducible in spirit, and almost completely misleading.

Mohinish S

SnackOnAI Blog

LFM2: Liquid AI's Hardware-in-the-Loop Architecture Search Found That SSMs Were Not the Answer for Edge Devices, and the Result Runs 2x Faster Than Qwen3 on a Galaxy S25

Aug 6, 2026

•

16 min read

LFM2: Liquid AI's Hardware-in-the-Loop Architecture Search Found That SSMs Were Not the Answer for Edge Devices, and the Result Runs 2x Faster Than Qwen3 on a Galaxy S25

Most efficient model architectures are designed on paper, benchmarked on A100 clusters, and then discovered to behave differently on the actual target hardware.

Mohinish S

SnackOnAI Blog

HeadRoom: The Context Compression Layer That Sits Between Your Agent and Your LLM Bill, and Gets 92% Token Reduction Without Touching Your Code

Aug 1, 2026

•

13 min read

HeadRoom: The Context Compression Layer That Sits Between Your Agent and Your LLM Bill, and Gets 92% Token Reduction Without Touching Your Code

Every coding agent you run is sending far more tokens to the LLM than the task requires. A code search returning 100 results generates 17,765 tokens. The LLM needs 1,408 of them to find the answer.

Mohinish S

SnackOnAI Blog

OpenWorker: Andrew Ng Built the AI Coworker That Delivers Finished Files, Not Chat, and Put the Agent Loop on Your Machine

Jul 31, 2026

•

11 min read

OpenWorker: Andrew Ng Built the AI Coworker That Delivers Finished Files, Not Chat, and Put the Agent Loop on Your Machine

Every AI assistant tool available today delivers the same output: a text response. You paste it somewhere, format it, send it, file it, or act on it yourself.

Mohinish S

SnackOnAI Blog

CAMEL: The First Multi-Agent Framework That Used Its Own Agents to Generate the Data That Trained Its Competitors

Jul 30, 2026

•

13 min read

CAMEL: The First Multi-Agent Framework That Used Its Own Agents to Generate the Data That Trained Its Competitors

The dominant narrative around multi-agent frameworks is that you pick one based on features: does it support tool calling, memory, RAG, streaming?

Mohinish S

SnackOnAI Blog

Colibri: A Single C File Runs a 744B MoE on 25GB of RAM by Treating Every Expert as Data to Be Staged, Not State to Be Held

Jul 29, 2026

•

14 min read

Colibri: A Single C File Runs a 744B MoE on 25GB of RAM by Treating Every Expert as Data to Be Staged, Not State to Be Held

The assumption every MoE inference system makes is wrong. It assumes the model must fit in memory.

Mohinish S

Mohinish S

The Postgres Predicate That Reads Your Table Twice

Jev Gives Up Writing And Gets A Type System

MetaGPT Wrote The Org Chart Into The Prompt, Then Quietly Retired Half Of It

BMAD Method Compiles Its Prompts, In Five Skills Of Thirty

Orca Supports Twenty Nine Agents With One Terminal

DeepSeek Harness Ships Hot Reload In One Profile Of Five

Obscura Runs Every Page Through One JavaScript Thread

llmfit's Fallback Estimate Can Be Fourteen Times Too Slow

OSWorld's Hardest Baseline To Beat Is An Agent That Gives Up

Model Optimizer's Own Benchmark Says Four Bit Is Slower

CUA-Lite's Parallelism Number Is Just The Memory Ratio

OpenSEO Is Very Good At Three Percent Of AI Citations

Anthropic's Alignment Agents Did Not Out-Think Anyone

Qwen3-VL And The Multimodal Tax

Rex-Omni, Or What Reinforcement Learning Actually Fixed

FreeToken, Or Why Your Laptop Was Never The Bottleneck

LinkedIn Benchmarked Nine Memory Systems For Its Hiring Agent, Seven Lost To Pasting Everything Into The Prompt

Autograd Is Slow Per Operation And Free Per FLOP, And That One Sentence Tells You When To Use It

The Seedance Paper Is a Scoreboard, Not a Blueprint

AnyDoc Is Not Faster Than Docling. It Is Solving A Different Problem, And Almost Nobody Noticed

LFM2: Liquid AI's Hardware-in-the-Loop Architecture Search Found That SSMs Were Not the Answer for Edge Devices, and the Result Runs 2x Faster Than Qwen3 on a Galaxy S25

HeadRoom: The Context Compression Layer That Sits Between Your Agent and Your LLM Bill, and Gets 92% Token Reduction Without Touching Your Code

OpenWorker: Andrew Ng Built the AI Coworker That Delivers Finished Files, Not Chat, and Put the Agent Loop on Your Machine

CAMEL: The First Multi-Agent Framework That Used Its Own Agents to Generate the Data That Trained Its Competitors

Colibri: A Single C File Runs a 744B MoE on 25GB of RAM by Treating Every Expert as Data to Be Staged, Not State to Be Held

Quick Links

Subscription

Socials