Amit Naik

I'm open sourcing Leviathan: an indexer that lets AI agents search a database of any size without reading it.

At 1M records it hands your agent 436 tokens instead of 107,000, and it finds the answer 99% of the time.

github.com/elstongun/levi…

Tencent just released an omni-modal parsing model on Hugging Face

One 5B model that turns documents, images, charts, geometry figures, audio and video into a single structured JSON.

You can now train your own Decision model with our free notebook! 💡

Qwen3.5-4B will generate decisions instead of text on just 8GB VRAM locally.

Learn to data prep (state, questions, gold answers), train, serve.

Notebook: colab.research.google.com/github/unsloth…
Guide: unsloth.ai/docs/basics/tr…

We’re open-sourcing Drex 1.5 today!

The smallest #1 decision-making model, built for local and cloud deployment with a 128k context window.👨‍💻

Compatible with Ollama, Unsloth, OpenCode or plug it into Claude, Codex, Cursor, Hermes, Claw and more.

Try it now: nace.ai/developers
Cloud users: Claim a $25 sign up bonus

Today at Google Cloud’s Gemini at Work event, we announced Gemini, a new single universal agent for work that has all of your business context and can be used for everything from knowledge work to answering questions, and content creation to coding, all from a single prompt box.

It’s built around core architectural principles:
☑️ Unified Agent: Gemini can answer questions, do knowledge work and generate code from a single prompt box.
☑️Access: Gemini is web-based and can be accessed from any device and integrated into third-party applications. It can also operate without a dedicated user interface.
☑️Persistent Execution: It runs in the cloud, which means it maintains a single set of memories, context and one personalization graph no matter where you access it.
☑️Multi-Agent Orchestration: Gemini can create sub-agents to tackle multi-step tasks and can also act as a coworker agent with a defined role and its own dedicated identity
☑️Context: Gemini knows your tools, data, and work history and learns how you work the more you use it.
☑️Model Choice Flexibility: It orchestrates across multiple models to deliver optimal quality and lower your costs.

.@Sonderling47: "I am hereby suspending from the Permanent Labor Certification Program some of the largest IT outsourcing firms in the world — Cognizant, Infosys, Tata, Wipro, HCL, and Capgemini — and also, due to multiple active federal investigations, @USDOL is suspending Microsoft and Adobe from the program."

How I’ve Used AI to Generate a Few Hundred Thousand Dollars in New Revenue in 4–6 Weeks

I have a story about how I have used AI coupled with changed human behavior to drive a significant change to business results in my business and very much increase the top line.

I am a vocal critic of useless masturbatory vibe coding and applications of AI which do not actually meaningfully change business outcomes in any way (If you’re new here)

But

I’m also simultaneously an avid user of AI, praying that somehow the gods of superintelligence will rescue my meager profit margins and save me from the permanent underclass.

Here’s something I’ve done over the past 4–6 weeks that has added a few hundred thousand dollars in top-line revenue to our business.

The Problem

My current cardinal sin as a business owner is that I allow proposals to stack up. I do site visits, and I’m unable to turn around a bid in a timely manner.

The way you ACTUALLY solve this problem is by simplifying your service offering as a business, so you can provide pricing on the spot and do a one-call close (aka sit down at the kitchen table and hammer out a contract with a client).

I’m dumb, and I’ve designed a business that is too complex to be able to provide that (at this time). As a side note, caveat: I’m working on rearchitecting everything so that we get there, but we’re not there right now.

At the moment, we’re living the experience that most contractors fit into:

1. You get a call.
2. Somebody says, “Hey, can you come out and look at this and give me a price for how much it would cost to do it?”
3. You go out, look at the job, take some measurements, and write down the scope.
4. You try and figure out how much it’ll cost, and then you go back to the client with the pricing.

This takes forever, is super inaccurate, is a horrific experience for everybody involved, and is the reason why it’s hard to get contractors to come out and it’s hard to get proposals out of them.

It’s not because they hate money or because they’re bad at their job. It’s just that trying to reinvent the wheel every single time you put a proposal together is an impossible exercise.

As the contractor or service provider, you’re terrified of providing the wrong price because, if you give the wrong price, it ends up being, at worst, a bad experience. It is, at best, a bad experience for the client, and at worst, it’ll put you out of business and into litigation or something of the sort.

So you get stuck on perfection, and as we all know, perfection is the enemy of good, and the bids pile up.

We get 5 to 7 leads a day, and there are times when I’m running 3 to 5 sales meetings per day. As you can imagine, if there’s any level of complexity to those jobs, it’s just an impossible exercise.

You end up having clients calling you every single day, texting, asking, “Where’s my proposal?” They think you’re ghosting them when, in reality, you’re just trying to do your best to somehow deliver projects while also pricing new projects and figuring out how to keep everything together at the seams.

Here’s how I’ve used AI (credit where credit is due) to go from perfection paralysis to a good-enough output, and how it’s driven multiple six figures of revenue for us here at West Coast Deck.

The Tools

I’m not using some sort of crazy technology stack. I’m not even using any of these MCP connectors or API connectors. I still don’t know what an MCP is. I don’t even know what an API is, so I don’t have the ability to figure any of that stuff out. Which means that if I can do this, you guys can probably figure it out too.

The tools I’m using are as follows:

* Quo w recording
* A Plaud Pin Pro
* A ChatGPT Pro subscription
* A PolyCam subscription
* A Company Cam subscription
* A Stanley Fat Max 30 ft. tape measure

Here’s What I Do

The first thing is, I capture all transcripts of phone conversations and email traffic that we’ve had with the client.

Next, I drive to the site. I park outside, and I turn on the plot plan. I start narrating. I say, “Here’s the address. I’m at 123 Main Street. This is Tom and Judy’s house. They have this type of deck project with this type of scope. I’m about to walk into the site and do a job walk and review the work.”

Then I start the meeting. I tell the client I’m using AI tools to record, and then I essentially do a very verbal job walk and interview with them.

I ask them a lot of questions about what they want to get done, and then I intentionally regurgitate back to them everything that they’ve said. I discuss my findings on the job site with them. I talk through different things. I pull measurements with my tape measure, say the measurements out loud, and talk through the universe of outcomes and options. I talk about materials. I try to be as verbal as possible, knowing that my AI recorder is capturing everything.

Now, here is part of why I think this has been so effective for our business.

The forcing function of me going through all of the options and being super verbal helps to create expectations with the client, help them understand what will happen, what might be possible, and what the general job scope is. It’s forcing me to be overly communicative with prospects.

Before I leave the job site, I take a Polycam 3D LiDAR scan of the entire backyard, which is useful for taking measurements and referencing later if I need to look at details.

I also then open Company Cam, and I do a video walkthrough where I reiterate everything that we just reviewed with the client on video. It’s like triPLICATE information: recording what scope is, what needs to happen, and what the next steps are.

After the Site Visit

The next step is:

1. I say goodbye to the client.
2. I go out to my truck, sit down, and talk to the Plod pin a little bit more about anything that wasn’t appropriate to discuss in front of the client or that maybe I missed and that I need to capture.
3. I end the recording.
4. I transfer the data from the Plod pin into the Plod app.
5. I copy the transcript from the recording into ChatGPT.

At that point, I’m also adding any previous transcripts from Quo, any pictures that they’ve sent us, any videos that they’ve sent us historically, and any additional information.

There are three more things that happen in the process:

1. I send a recap email to the client with next steps.
2. I generate a scope of work and proposal.
3. I reference our historic library and universe of proposals for pricing.

My Super Scientific AI Prompts

Normally, you have to buy my course to get my super scientific AI prompts, but I’ll give them away for free here.

The first prompt is this:

Please write a site visit recap email based on these transcripts and information, and send it to the client.

Took me months and months to come up with that prompt. Should make you millions of dollars

Even just that little step has made a huge difference in the customer experience. Literally 5 minutes after the site visit, they get a super detailed email from me recapping everything, talking about next steps, and what the project experience will be like for them.

The next prompt is also probably worth millions of dollars. Please feel free to steal it.

Please write a construction proposal based on the site visit transcript and all the information. Reference my previous proposals.

And it spits out something that honestly is plenty good enough to sit down and review as a client.

The next step is: you can ask it to reference pricing from previous proposals, or you can just look at your price book and manually input pricing into the detailed proposal.

The Results

I can’t wait for all the comments from people saying, “Oh, you should do this better.” That’s part of why I post this kind of stuff, because I want people to tell me how to do it better.

We already have a V2, V3, and V4 of this planned, but this has first concept already resulted in a couple hundred thousand dollars in initial sales in just 4 to 6 weeks.

Anybody could figure this out. It does not take any sort of AI wizardry. It’s just basic, rudimentary tools that are available to anybody.

Provecho! (Happy hunting)

A new personal AI assistant has arrived. Meet Tab, which announced Tuesday that it had officially emerged from stealth.
spr.ly/6015BG2WZd

Most AI is a solo act. We’re making it a team sport.

Today at Team ‘26 Europe, we’re introducing Agentic Multiplayer Protocol (AMP).

How it works:
🤝 Work alongside your agents across every surface.
🤝 Know who’s who, who’s doing what, and who did what, when.
🤝 Agents are grounded in organizational context to deliver real impact.

It is very cool to see llama.cpp on the big stage in todays Windows event!

The software and hardware stacks are finally coming together. Our community has put a lot of hard work in the past years and it shows.

Looking forward to more users embracing local AI.

Today we release Open d1: two open-weight multimodal models in our d1 decision model family.

> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.

1/

Introducing 𝗩𝗲𝗹𝗮 𝟮.𝟬: Our newest 𝗢𝗽𝗲𝗻 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗥𝗼𝘂𝘁𝗶𝗻𝗴 𝗠𝗼𝗱𝗲𝗹𝘀, bringing 𝘀𝗽𝗮𝗻-𝗹𝗲𝘃𝗲𝗹 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀 to System One.

HF Collection 🤗:
huggingface.co/collections/vl…

#OpenSource #AI #LLM #vLLM #HuggingFace

Today we’re announcing Personal Agent Protocol — an open standard @Meta and @SierraPlatform are developing along with industry partners at @Genesys, @instinct, @RocketOTD, @Shopify, @stripe, and @Walmart. It will help define how personal agents interact with businesses and is open for anyone to implement. You can read more here - and if anyone is interested in joining let me know! sierra.ai/blog/introduci…

Meet Mistral Large 4, aka Le Chonk.

• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.

Open weights release end of October.

We are making Markdown files work natively across Google Drive and Docs.

You can now preview .md files in Drive, open them in Docs, edit them, comment on them, and collaborate without converting the file into a Google Doc!

Our (@cdatasoftware) model benchmark is reproducible on your own machine. Copy three Google Sheets, clone the harness, and do a dry run before spending anything on models. Then get Connect AI to see how much of accuracy, safety, and cost comes from the model, and how much from the gateway beneath it.

Guide: bit.ly/4dc2Rl2

We turned Claude Code, Codex, Hermes, Pi, @opencode and other coding harnesses into RL environments. No changes to the harnesses, no changes to the training code. Any open model, any task set, fully open source my friends!

Same model, same weights: 62% under Mini-SWE-Agent, 33% under Claude Code. But training inside a real harness normally means reimplementing it as an environment, so most models get trained in a scaffold nobody actually ships.

The fix is a proxy, not a rewrite. The harness thinks it's talking to a model API. It's actually talking to a capture proxy that speaks the 4 formats coding agents use (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Gemini), forwards to @vllm_project, records the exact token IDs and logprobs vLLM sampled, and hands TRL sequences it can train on. The harness becomes the environment. 10 harnesses run through it today, none modified.

And because you control the reward, you can shape behavior the harness never asked for. We added a small bonus for solving a task in fewer tool calls: on tasks it already solved, the model now uses 31% fewer calls, in every harness, and about half under Codex.

Tested on LFM2.5-2.6B from @liquidai:
→ Train in one harness: better mostly in that harness (OpenCode 34% → 58%).
→ Train in 4 at once: better in all 4 (42% → 54%).
→ SFT on 3,189 rollouts from Qwen3.8-27B instead: plateaus at 47.5%, below both RL runs.

Everything is open and reproducible: the capture proxy in OpenEnv, the trainer in TRL, the tasks, the SFT data, the training code and all 7 trained models. Bigger models and bigger runs next.

Full guide: huggingface.co/spaces/FineEnv…

Small bird, fast wings, Kolibri is here.

78B parameters. 3.46B active. Up to 1M tokens of context. Built in Europe.

Now the weights are yours. Run it on your own hardware, under Apache 2.0.

Introducing Underdog, your Private Personal AI on devices you already own

Today we're announcing our backing from @a16z @khoslaventures @HummingbirdVC Anthology (@AnthropicAI @MenloVentures) @patrickc @naval @rauchg @polynoamial @Thom_Wolf @Mascobot @tszzl @OfficialLoganK and other top AI leaders

Our mission is to provide free, capable, reliable and private AI to billions of people.

Underdog’s Law: today’s frontier intelligence reaches your devices in six months.

We’re starting with fast, capable AI that runs on consumer hardware. By co-designing models and inference engines, we’re pushing the frontier of capability, speed, power and data efficiency.

Our research also spans agentic commerce, confidential inference, and how AI will reshape the internet economy.

Privacy and capability no longer needs to be a tradeoff. If you believe in this future, join us. Time to build.