dottxt is focused on developing tools to improve the integration of language learning models (LLMs) with software systems, aiming to enhance the reliability and efficiency of AI applications.

dottxt

.txt | 2,757 followers on LinkedIn. Making AI speak computers | We provide a fast, reliable model that generates valid JSON 100% of the time via an API. For information extraction, classification, agents workflows, synthetic data generation.

Jeffrey Enos

Brandon T. Willard

Software Engineer at .txt

Aaron Schwartz

Staff DevOps Engineer at .txt

Dan Palmer-Bancel

.txt

The world of AI is becoming a "Smol world after all"! Our friend's at Hugging Face recently released a new series of models: SmolLM2! The aim of these models is to provide lots of power that can be run on portable or other low memory devices. We decided it would be fun to demo the power of these Smol models when combined with structured generation using Outlines! Presenting the "Bunny B1" - a demonstration of a natural language interface for calling applications on a mobile device. The model is provided with a list of possible apps that can be called and uses the SmolLM2 model along with Outlines to transform that natural language request into an application call. The this demo version contains 4 applications that can be called: - Food delivery (order_food) - Ride share app (order_ride) - Getting the weather (get_weather) - Text messaging (send_text) Of course, in the demo, all we have is the function that would be called, but it's not difficult to imagine how this interface could be hooked up to real devices and apps. Because the SmolLM2 models are so small, and Outlines so fast, tools like this can be incorporated into a wide range of devices! Best of all: all of this is open source! So if you want you can take this demo and extend it yourself!

New Rules for AI There are hidden rules in every system. Rules to govern the sentences we write, the software we code, and the music we play. These rules aren’t created to subjugate, but to integrate. Without them, the different parts of a system can’t work together to form something greater. Some believe that these rules reflect our unique way of thinking and conceiving the world. But for computers alike, without structure there is chaos. Keywords and variable names, out of order, don’t make a program. You need a grammar — a syntax — to make the system work. For almost a century, we’ve been teaching our machines the grammars of our world to help us work, play, and create better. But as of yet, our AIs remain stubbornly resistant to learning them. While they can string together sentences to emulate human syntax, they’ve yet to learn the syntax of computer, and of the applications that run our world. Without more predictable outputs, LLMs simply can’t be integrated into our systems and businesses. It’s their fundamental flaw. We believe that this demands a new programming paradigm, which allows LLMs to respond in a way that software understands. Until then, our generations are fun. But they are not transformative. Until then, it’s just the beginning for how AI will change our world. 👉 Our mission is to help AI speak the language of every application. This is why we built .txt. A tool to make LLMs reliable enough for the world to build on. A spellcheck for programming the future. And more broadly, an ecosystem where developers can design, execute, deploy, and evaluate LLM applications. But how? LLMs are probabilistic by nature, so we believe the only solution is implementing guard rails at the outset based on your unique criteria… taking all potential probabilities and setting all the useless ones to zero. In other words, .txt stops LLMs from making errors before they even make them. Fewer tokens. Less money. More efficiency. We are not the average AI engineers. We cut our teeth in applied statistical modeling — a job where you make bespoke models specially designed to give only correct answers. To us, applying this field to LLMs is the logical next step. The missing layer that will make AI truly useful, cutting out the noise and uncertainty endemic to LLMs. It’s the thing that makes LLMs act like computers. Finally. If your goal is to sound conversational and believable enough, the everyday language works just fine. But to be good enough to compute with, the level of clarity must be far higher. In this sense, LLMs are, ironically, a regression! So, let’s create the new rules for AI and invent a new way to program with them… together. Join the discussion on Discord. Or just tweet at us. We can’t wait to see what you create!

We just released outlines-core 0.1.0 with the .txt team 🙌 It's a Rust port of the core parts of outline's structured generation algorithms. Benefits: - faster compilation - lightweight library - easier to create language bindings links below 👇

Erik Kaunismäki

These 39 fast-growing startups each raised $10-50M from top-tier funds in recent weeks and are now HIRING: 1) OpenGradient - open-source model hosting (US remote) 2) Lightdash - self-serve BI (remote) 3) .txt - build on LLM applications (remote) 4) Qodo (formerly Codium) - quality-first AI tools (Tel Aviv / US) 5) Tabs - b2b revenue automation (NYC) 6) Valdera - chemical sourcing (US remote) 7) Lanai Software - AI model integration (Bay Area) 8) Alchemy - on-site pharmacy management (remote / Miami / San Antonio) 9) Tebi - integrated hospitality management (Amsterdam) 10) Decagon - gen AI customer support (Bay Area) 11) Braintrust - enterprise AI development (Bay Area / remote) 12) Distributional - AI system testing (US / Canada remote) 13) RadiantGraph - personalized healthcare AI (Bay Area / remote) 14) General Galactic - renewable synthetic fuel production (El Segundo / LA) 15) Salient Motion - advanced electric propulsion (Irvine) 16) Tennr - healthcare document reasoning model (New York) 17) Fern (YC W23) - API SDK generation (NYC) 18) Omnea - Automated purchasing approvals (Europe / NYC) 19) Daylight - Web3 wallet functionality (NYC / remote) 20) autone - AI retail inventory management (London / NYC / Paris) 21) AirOps - LLM-powered workflow creation (NYC / Bay Area) 22) Granular Energy - clean energy trading software (Europe / NYC) 23) Claim - privacy-focused cashback shopping (US) 24) Rogo - Generative AI for financial firms (NYC) 25) Kick - AI-assisted bookkeeping (US Remote) 26) Cofactr - electronic parts management (NYC / US remote) 27) Numeric - AI-powered accounting automation (NYC / Bay Area) 28) LiveFlow - Automated financial reporting (NYC) 29) Port IO - internal developer portal (Tel Aviv / London / Boston) 30) Galileo 🔭 - Evaluation intelligence for AI Teams (Bay Area) 31) Ghost - B2B marketplace (LA / US remote) 32) Noetica AI - software for benchmarking deal terms (NYC) 33) Relyance AI - data governance platform (Bay Area / remote) 34) Logik.io - AI-powered CPQ (Deerfield / Bangalore) 35) Centaur Labs - health data annotation (remote US) 36) Wispr Flow - voice dictation (Bay Area) 37) Pallet - Modern logistics (Bay Area / India) 38) Mstack Chemicals - specialty chemicals (Remote US) 39) Resolve AI - AI production engineer (Bay Area) Adding direct links to their career pages in the comments 👋 follow me for more lists like this every week: Ben Lang ✨ thank you to Harmonic (startup database) for helping me source the list

Ben Lang

As an AI-Engineer you often start to feel like you're "prompt-and-praying" more than engineering. If you've worked with AI long enough, you've definitely been there: You have an idea for a way to fix an issue with your LLM, you draft up a new prompt, that prompt works on a few examples, then you run evaluations (hopefully) and... it's worse! One of my favorite parts of structured generation with Outlines is that it allows you to feel like an engineer again. My favorite part of the pattern? Just like with machine learning you start with *data*. Here's the flow: 1. Take real examples of the output you'd like to see from the LLM. 2. Quickly draft up the structure you think is right (maybe JSON or a regex) 3. Before you start generating anything, make sure your proposed structured *matches* your data! 4. Then generate a few examples using Outlines and your favorite open LLM. 5. Inspect the output, see a bug: goto step 2! With this process your LLM code gets iteratively and predictably better!

There are hidden rules in every system. Rules to govern the sentences we write, the software we code, and the music we play. These rules aren’t created to subjugate, but to integrate. Without them, the different parts of a system can’t work together to form something greater. Some believe that these rules reflect our unique way of thinking and conceiving the world. But for computers alike, without structure there is chaos. Keywords and variable names, out of order, don’t make a program. You need a grammar — a syntax — to make the system work. For almost a century, we’ve been teaching our machines the grammars of our world to help us work, play, and create better. But as of yet, our AIs remain stubbornly resistant to learning them. While they can string together sentences to emulate human syntax, they’ve yet to learn the syntax of computer, and of the applications that run our world. Without more predictable outputs, LLMs simply can’t be integrated into our systems and businesses. It’s their fundamental flaw. We believe that this demands a new programming paradigm, which allows LLMs to respond in a way that software understands. Until then, our generations are fun. But they are not transformative. Until then, it’s just the beginning for how AI will change our world. Our mission is to help AI speak the language of every application This is why we built .txt. A tool to make LLMs reliable enough for the world to build on. A spellcheck for programming the future. And more broadly, an ecosystem where developers can design, execute, deploy, and evaluate LLM applications. But how? LLMs are probabilistic by nature, so we believe the only solution is implementing guard rails at the outset based on your unique criteria… taking all potential probabilities and setting all the useless ones to zero. In other words, .txt stops LLMs from making errors before they even make them. Fewer tokens. Less money. More efficiency. We are not the average AI engineers. We cut our teeth in applied statistical modeling — a job where you make bespoke models specially designed to give only correct answers. To us, applying this field to LLMs is the logical next step. The missing layer that will make AI truly useful, cutting out the noise and uncertainty endemic to LLMs. It’s the thing that makes LLMs act like computers. Finally. If your goal is to sound conversational and believable enough, the everyday language works just fine. But to be good enough to compute with, the level of clarity must be far higher. In this sense, LLMs are, ironically, a regression! So, let’s create the new rules for AI and invent a new way to program with them… together. Join the discussion on Discord. Or just post at us. We can’t wait to see what you create! -Remi, Brandon, and Dan

We are proud to announce our $11.9M fundraise to help solve AI's fundamental flaw. The two successive rounds were led by Elaia and EQT Ventures with the participation of Seedcamp , Common Magic , Kima Ventures , Script Capital , STATION F and many amazing angels investors Roxanne VARZA , Julien Chaumond , Bob van Luijt , Erik Bernhardsson , Jean-Louis Quéguiner , Rick Lamers , Pierre Valade and many others. While AIs can string together sentences to emulate human syntax, they’ve yet to learn the syntax of computers… and of the applications that run our world. Without more predictable outputs, LLMs simply can’t be integrated into our systems and businesses. We believe that this demands a new programming language, which allows LLMs to respond in a way that software understands. So, .txt is a way to make LLMs reliable enough for the world to build on. A spellcheck for programming the future. And more broadly… an ecosystem where developers can design, execute, deploy, and evaluate LLM applications. We cut our teeth in applied statistical modeling — a job where you make bespoke models specially designed to give only correct answers. To us, applying this field to LLMs is the logical next step. The missing layer that will make AI truly useful, cutting out the noise and uncertainty endemic to LLMs. It’s the thing that makes LLMs act like… computers. Finally. So, give it a try. Join the discussion on Discord. Or just tweet at us. We can’t wait to see what you create! Remi, Brandon, and Dan

DOTT

CO:DOTTXT

dottxt