here’s how we spent 1T tokens at ramp
200 tokens - receipt processing
350 tokens - invoices
790 tokens - reimbursements
999,999,998,660 tokens - accidentally pushed the api key to github
94
182
7,688
494,899
Anthropic's AI Engineer source code is fully public / there is no server
there is no separate backend. they just use the same api.anthropic.com/v1/message… api in a loop with tool use
all packaged into a single file: gist.githubusercontent.com/1…
tools available:
1. dispatch_agent - Creates a specialized agent for searching files and code with access to GlobTool, GrepTool, LS, View, and ReadNotebook tools.
2. Bash - Executes bash commands in a persistent shell session. Can't use search commands like find/grep or read tools like cat/ls.
3. GlobTool - Fast file pattern matching using glob patterns like "**/*.js".
4. GrepTool - Searches file contents using regular expressions.
5. LS - Lists files and directories in a given absolute path.
6. View - Reads files from the filesystem (up to 2000 lines).
7. Edit - Edits parts of files by replacing text with new text.
8. Replace - Overwrites entire files with new content.
9. ReadNotebook - Reads Jupyter notebook files (.ipynb).
10. NotebookEditCell - Edits specific cells in Jupyter notebooks.
11. StickerRequest - Displays a shipping form for Anthropic/Claude stickers.
Introducing Claude 3.7 Sonnet: our most intelligent model to date. It's a hybrid reasoning model, producing near-instant responses or extended, step-by-step thinking.
One model, two ways to think.
We’re also releasing an agentic coding tool: Claude Code.
24
129
1,648
368,014
Problem: Getting LLMs to output valid JSON in the format you want is hard
Solution: ONLY generate values, feed model with keys and JSON structure. Constrain outputs with custom sampling
New project: Jsonformer: A Bulletproof Way to Generate Structured JSON from Language Models!
41
130
1,238
378,956
got @anthropicai Claude Code working with OpenAI models lol
i set up an proxy server that mimics the anthropic /v1/messages api, forwards requests to OpenAI
maps:
- Sonnet 3.7 -> 4o
- Haiku 3.5 -> 4o-mini
Sonnet3.7 is still better than gpt-4o at agentic and coding tasks, able to run longer sessions, follow instructions more closely, and bring tasks to completion. Will be testing 4.5 and o3-mini today
source below:
59
46
1,029
147,865
added gemini 2.5 pro support to Claude Code
feels faster and smarter than Sonnet 3.7
my go to local coding assistant now
link below ⬇️
38
42
892
97,745
at @tryramp we use LLMs to find the 5 most valuable mins of audio from the 1000+ customer calls we make every day
narrated by TTS + compiled into a 5 min podcast sent to the entire team
as a product owner it'd be nice to have an llm summary of everything my users did yesterday
calling out cool success stories or troublesome error states i should reach out to debug
has anyone tried such a thing? i am thinking about prototyping it with public val town data
22
31
539
237,894
🤔 extracted the full ~5000 token claude3.5sonnet claude.ai system prompt: gist.github.com/1rgs/b31a1de…
this is a great template for function calling / tool use
notes:
artifacts: seem to be a fully in-context abstraction, model not finetuned for it
allowed types: markdown, html, svg, react (tailwind, lucide-react, recharts, shadcn/ui), mermaidjs.
8 fewshot examples, all types + example of not using an artifact
good artifacts are >15 lines, modifiable, self-contained, external use
avoid artifacts: simple, explanatory, conversational, context-dependent
one artifact per message unless requested
prefer in-line content
artifact steps: think in <ant_thinking>, wrap in <ant_artifact> with identifier, title, type
artifact types: application/vnd.ant.code (code, specify language), text/markdown, text/html, image/svg+xml, application/vnd.ant.mermaid, application/vnd.ant.react
complete content, no truncation
err on not creating artifact if unsure
claude: created by anthropic, current date, 2024, knowledge updated april 2024
claude: no urls or videos, provides info regardless of views, sensitive topics carefully, helps with analysis, coding, writing, teaching, discussion, uses markdown for code
claude: face-blind in images, describes without identifying humans
claude 3 family: haiku (fast), opus (writing/complex tasks), 3.5 sonnet (most intelligent)
claude: thorough for complex, concise for simple, responds in user's language
full prompt: gist.github.com/1rgs/b31a1de…
14
54
412
53,291
introducing 🪄genweb: the first software 2.0 web framework 🪄
genweb is a new way of building web apps: instead of a frontend and backend codebase, an LLM is the backend and the frontend
it interprets user actions and dynamically generates UI in real-time
welcome to the simulation genweb.rahul.gs/app?appId=e3…
jensen said apps will be generated not rendered, this is the start
100% Fully Software 2.0 computer. Just a single neural net and no classical software at all. Device inputs (audio video, touch etc) directly feed into a neural net, the outputs of it directly display as audio/video on speaker/screen, that’s it.
20
23
359
100,523
we're hiring full stack engineers to work on llms at ramp
if interested, dm me with examples of real things you've built
come work on real deployments and learn how to drive enterprise value
we're a small and mighty team
what we've worked on in the last year:
- multi-step agents for document extraction / ocr (sota accuracy, probably). llm agents + constraint solvers
- low-latency next action prediction in our web app (more soon)
- ramp tour guide: nitter.app/tryramp/status/1792659…
- web agents for solving c*ptchas
- codebase import cycle removal with ast parsing/graph cutting algorithms + llms in our python monolith backend
- sales outbound automation and lead scoring agents
- llm model routing between third party providers (+per feature cost tracking)
- llm infra: embedding/reranking/generation finetuning and on-prem deployment/inference
- structured extraction (github.com/1rgs/jsonformer)
- customer feedback extraction from meeting recordings / routing to marketing
- internal tools for: underwriting team/product team/sales/customer support teams
- global search / function calling copilot
- receipt matching (retrieval)
- sms llm interface (function calling)
- suggested memos
- automated accounting coding
- natural language report generation
+ more
Introducing Ramp Tour Guide: an AI Agent that can show you how to do anything on Ramp!
Today, we'd like to share a sneak peek of Ramp's near future. As Ramp grows in functionality, we want to make all of it easily accessible to all of our customers. To do that, we're demoing a first-of-its-kind AI agent that can show and tell people how to accomplish anything with Ramp.
The Ramp Tour Guide knows Ramp inside and out. You can ask it how to do something on our platform and it'll walk you through every step of the way.
18
20
326
82,496
🎉 new project: Clarity! A reading app that offers a fresh approach to consuming text. Instead of the traditional linear reading style, Clarity allows you to read depth-first, diving into the details that interest you most.
15
26
304
56,625
claude.ai generates CoT tokens within <antThinking> tags, hidden from user on the server
16
11
225
53,895
Excited to announce that we're joining forces with one of our customers, @tryramp, where we will help build the future of AI + finance
Exclusive: @tryramp makes its 2nd acquisition, scooping up Cohere.io, which has built out an AI-powered customer support tool.
techcrunch.com/2023/06/26/as…
12
3
150
37,966
I kept having to debug prompt issues with open models
So I built OpenAI's Tokenizer page for all tokenizers on HuggingFace: Llama, Mistral, GPT2, MPT, Persimmon, T5 etc
github.com/1rgs/tokenwiz
check it out here: tokenwiz.rahul.gs
8
22
141
21,536
jsonformer + openai!
openai.com/index/introducing…
"deterministic, engineering-based approach to constrain the model’s outputs to achieve 100% reliability"
Problem: Getting LLMs to output valid JSON in the format you want is hard
Solution: ONLY generate values, feed model with keys and JSON structure. Constrain outputs with custom sampling
New project: Jsonformer: A Bulletproof Way to Generate Structured JSON from Language Models!
3
8
143
26,776
at @sequoia ai ascent last week
i spoke about why ai agents, even from companies like @microsoft @apple @google, fail
+ how to solve it with a simple fix
10
3
142
33,924
My favorite thing to do on Modal - running massively parallel GPU finetune jobs
At Ramp, we’ve trained hundreds of LLMs *at the same time* without the infra hassle - Modal allows us to move insanely fast
(1/2)
Modal is generally available today, and we also raised a Series A!
modal.com/blog/general-avail…
3
6
123
53,757
try it here: github.com/1rgs/claude-code-…
run with: `ANTHROPIC_BASE_URL=http://localhost:8082 claude`
9
6
129
7,848
literally free productivity, most people one shot themselves by working or sleeping in rooms with high levels of CO2
buy a monitor that tracks levels over time so you can see how high it went when you were asleep, aranet makes a good one
I asked Erik @bernhardsson why high CO2 levels in your office are such a big deal:
"I'm not a health nut. But one of the things I've been radicalized on is CO2 levels.
There's a real relationship between CO2 levels, productivity, and cognitive performance. And CO2 levels are usually way too high in offices and schools.
Normal air CO2 levels are 300-500 PPM. In offices it often hits 1,000 or 2,000. Airplanes can get up to 2,500. You start getting brain damage at 5,000.
So I went and bought a bunch of CO2 monitors for the office. And I look at them every day. And I open windows anytime we get too high."
5
12
128
55,722
Looking at @anthropic-ai/claude-code source from NPM
there's a sticker tool! ask claude code for free anthropic stickers
Introducing Claude 3.7 Sonnet: our most intelligent model to date. It's a hybrid reasoning model, producing near-instant responses or extended, step-by-step thinking.
One model, two ways to think.
We’re also releasing an agentic coding tool: Claude Code.
2
2
110
11,762
Shoutout to Ramp engineer Andrew Gu who was on the coaching staff.
Congratulations to Team USA for winning IMO 2024!
4
89
16,449
so pumped to share the news, and ...we are just getting started! how we got here ↓
(thanks @sarahintampa for the awesome coverage!)
techcrunch.com/2021/03/17/co…
5
6
78
thanks to modal sandboxes, everybody gets a free vm on os.rahul.gs
each tab is it's own FULL VM running ubuntu
dont mine crypto pls
It's true – @modal has raised a $87M Series B at a $1.1B valuation to advance the future of AI infrastructure.
Thank you to @Lux_Capital, @Redpoint, @AmplifyPartners, and others.
Now more than ever, AI demands a complete reinvention of traditional compute infrastructure
7
4
81
19,935
with the o1 release, reminder that claude.ai has been using thinking tokens for several months now
openai.com/index/introducing…
2
2
68
9,874
Jsonformer supports a subset of JSON Schema, including number, boolean, string, array, and object types. It's built on top of the HuggingFace transformers library, making it compatible with any model that supports the HuggingFace interface.
Try it — github.com/1rgs/jsonformer
4
7
63
7,562
New complex schema generation example live
With just a tiny 3b model (databricks/dolly-v2-3b)
github.com/1rgs/jsonformer/b…
2
5
55
6,411
Introducing our Vendor Search Tool!
No more generic SEO lists. No more fake reviews. No more wasting time on a bunch of different sites.
Just real, detailed information sourced from all across the internet — all in one place, all powered by Ramp Intelligence.
Find the right vendors for your business:
buy.ramp.com
Finding the best vendors used to be hard. Not anymore.
Introducing Ramp’s Vendor Search Tool.
See pricing, compliance, and growth trends, all in one place.
Powered by Ramp Intelligence.
Find the best software for your business: buy.ramp.com
3
1
51
7,015
i achieved 100% accuracy on 0.007% of swe bench
Sweep achieves 15.7% on SWE-bench!
Hi everyone, we’re building Sweep, an open-source AI developer that handles the easiest 30% of software tasks.
We’re thrilled to announce our results on SWE-Bench! We evaluated Sweep on a random 10% subset of the data.
Sweep correctly completed 15.7% of issues (1.9% more than Devin)!
3
48
9,279
Generating JSON is probably a common enough use case that hosted model providers should probably support an JSON only API
thoughts? @gdb @aidangomezzz @AnthropicAI
3
1
47
6,896
I finetuned an LLM on all my iMessages, try it on yours!
releasing code with sql queries, data processing, finetuning with PEFT and a chat CLI
github.com/1rgs/MeGPT
3
1
47
7,331
✅ agent that works
✅ general availability
✅ a real demo
We built a B2B SaaS sales company and here’s what it taught us about B2B SaaS sales 🧵👇 (but actually)
Today we’re launching Rox, the first publicly available AI agent swarm for the top sales teams, and in the private beta it already helped reps grow their books 30%.
2025 is going to be a huge year for growth. Enterprises are doubling next year’s revenue goals, but no one is doubling team sizes. Every rep will have to bring in more, and AI can help them do that. But there’s a wrong way and a right way.
Much of today’s AI aims to replace low-value work. But sales follows a power law: 90% of revenue comes from the top 15% of enterprise sales reps. The greatest gains will come from supercharging the highest value work — raising the ceiling, not the floor.
Rox equips the very best with a swarm of AI agents, acting as an army of analysts to help them plan, prioritize, research, engage, and keep up with their customers.
Over 35 of the best-performing enterprise sales teams have adopted Rox virally. For example, Ramp has rolled it out to their AE and AM teams, and we are now integrating their internal data systems with Rox. The Enterprise AE team alone gains 225+ hours per week to boost pipeline execution activities.
Rox is now in public beta. No barriers. No need to request a demo. Try it now for free → rox.com
5
1
43
4,561
openai: with structured mode vs without
in my benchmark, structured extraction mode is 13% slower, samples about the same number of tokens
code: gist.github.com/1rgs/4790c32…
6
5
39
4,031
huge
❓How to get models to generate structured output?
JSONFormer (by @rahulgs) and RELLM (by @mattrickard) are two novel approaches for this, now with (experimental) integrations to LangChain
JSONFormer Integration python.langchain.com/en/late…
RELLM Integration python.langchain.com/en/late…
5
40
11,413
Problem:
Generating structured JSON from language models is challenging. Current approaches like prompt engineering, fine-tuning, and post-processing often fail to produce syntactically correct JSON.
2
36
8,361
when i was at superhuman it would bother me immensely when people called us SuperHuman
we've come full circle
3
38
Solution: Jsonformer: A wrapper around HuggingFace models that only generates content tokens and fills in fixed tokens during the process. This makes it more efficient and bulletproof than existing methods
2
33
7,618
Jack Ma 5 years ago: “hate that AI is called Artificial Intelligence, I call it Alibaba Intelligence”
@elonmusk: “damn might end up being true”
1
2
31
9,021
Source: github.com/1rgs/clarity-read…
Based off of @andy_matuschak's amazing Evergreen notes and @OpenAI's "Recursively Summarizing Books with Human Feedback" (arxiv.org/abs/2109.10862)
And huge thanks to @yunyu_l @thesephist for feedback
2
32
1,808
literally everyone under the age of 25 who has invested in Cohere has asked if they can Venmo me the money
4
33
when I was making this graphic in Figma @yunyu_l told me to change to the latex font so it looks more academic
3
30
3,515
this is just the beginning, excited to be a supporter
Today we're excited to introduce Devin, the first AI software engineer.
Devin is the new state-of-the-art on the SWE-Bench coding benchmark, has successfully passed practical engineering interviews from leading AI companies, and has even completed real jobs on Upwork.
Devin is an autonomous agent that solves engineering tasks through the use of its own shell, code editor, and web browser.
When evaluated on the SWE-Bench benchmark, which asks an AI to resolve GitHub issues found in real-world open-source projects, Devin correctly resolves 13.86% of the issues unassisted, far exceeding the previous state-of-the-art model performance of 1.96% unassisted and 4.80% assisted.
Check out what Devin can do in the thread below.
29
5,195
🤯 our most requested feature is out!
Announcing Cohere Voice! The same frictionless experience, now with audio and video.
It’s that easy. 🎤📹
cohere.so/voice
28
was an honor to drop some hot takes on stage @aiDotEngineer
thanks for having me @swyx!
THE BITTER LESSON APPLIED TO AGENTS
(aka how to not be steamrolled by GPTNext)
Ramp just hit a $13b valuation and "every surface of Ramp is infused with AI"
TL;DR of @rahulgs' very well constructed @aidotengineer talk
as a syllogism
1. systems that scale with compute beat systems that don't
2. you should build systems such that they improve with more compute
3. exponentials are rare: when you find one, -actually- hop on for the ride (instead of subconsciously fighting it out of habit/fear)
4. therefore allow the agent to flex tools and self augment/improve rather than constrain it
has everything: single message, real life usecase from a major company, and LIVE DEMO (on conference wifi lol)
do not miss
1
1
27
3,783
didn't get access to copilot x yet so I wrote my own with gpt4
try bropilot, a rust cli that helps you write terminal commands
github.com/1rgs/bropilot/tre…
2
27
2,234
“In 15 words: deep learning worked, got predictably better with scale, and we dedicated increasing resources to it.”
- Gandhi
one of the hardest of parts of building a good agentic UX is integrating to the user's context. automation only works when we can make intelligent decisions without requiring a user to put in extra work.
proud to have led this project alongside many others at @tryramp 🤝
2
27
3,637
agree, 100% a mistake
Klarna using AI to rip out Salesforce and Workday is pretty magical at first glance....
but I've also seen this before:
- company sees 7-fig Datadog bill
- kicks off internal build to "save millions of dollars!"
- staffs up team of eng
- 6 months later, realizes their mistake
🧵
2
24
3,829
every visit to a genweb app goes straight to an llm, which renders the initial page in html
all “code” is in natural language, which is “interpreted” by an llm real time
user interactions are piped back into llm, which "rerenders" the page
every user session is a multi-turn LLM conversation. here's an example:
1
24
3,300
next few years are going to be crazy
if you're curious how many tokens are in your codebase: github.com/1rgs/token-trekke…
A bunch of our repos fit in one context window 🤯
Introducing 100K Context Windows! We’ve expanded Claude’s context window to 100,000 tokens of text, corresponding to around 75K words. Submit hundreds of pages of materials for Claude to digest and analyze. Conversations with Claude can go on for hours or days.
1
1
23
3,657
unlike traditional AI code generation (eg copilot, chatgpt, claude artifacts, devin), which outputs code, genweb is the LLM itself
llm -> code -> app ❌
llm -> app ✅
no js, no backend code - just natural language instructions and an LLM that simulates it
1
23
6,395
With Chime, we're bringing the magic of Cohere's seamless customer interaction tools to sales and marketing teams — super excited to get this out
1
23
genweb is a proof of concept for now, but with faster models and cheaper inference, this could soon be how all software is made
software 2.0 apps are malleable and squishy, not rigid and rules-based like it is today
(1) not every feature needs to be described, and the model fills in the gaps with “common sense”
(2) every user gets their own custom ui, tailored to their attributes, even with the same “source code”
here’s a playground to build your own genweb apps:
genweb.rahul.gs/
github.com/1rgs/genweb
1
23
2,237
was able to get access without getting off the waitlist:
copilot-workspace.githubnext… /<owner>/<repo>?task=<description>
What started out as an autocomplete pair programmer is now redefining the developer experience itself. Welcome to @GitHub Copilot Workspace: The Copilot-native developer environment — a place for all to create with code instantly in natural language.
github.blog/2024-04-29-githu…
1
21
6,885
from interviewing me for my first ever job at Superhuman to writing our first check at Cohere, Vivek has been a great mentor/supporter 🙏 thank you @vsodera — wouldn't be here w/o you
Proud to be one of the first investors in @CohereHQ. Their pixel-perfect screensharing experience is 🤯!
If you're a head of support, customer success, QA, onboardings, or sales, and want to use Cohere at your company, DM me.
/cc @yunyu_l @rahulgs @jasonhfwang
21
4.1 is the default model on Claude Code Proxy now github.com/1rgs/claude-code-…
feedback coming soon
1
21
1,836
vc: what’s ur mrr
me: haha we’re not sharing rn
vc: haha how many customers do u have and what is the average deal size
4
21
"LLMs are not enough" is going to age extremely poorly
from @leopoldasch
3
2
19
5,869