Mike Ritchie

2.8K posts

Mike Ritchie banner
Mike Ritchie

Mike Ritchie

@thisritchie

Founder @definiteapp (a data team that never sleeps)

Philadelphia, PA Katılım Ocak 2018
293 Takip Edilen1.2K Takipçiler
Sabitlenmiş Tweet
Mike Ritchie
Mike Ritchie@thisritchie·
The average company buys Snowflake, then Fivetran, then a BI tool, then spends 6 months connecting them. We spin it all up in 30 seconds.
English
0
1
10
878
Gergely Orosz
Gergely Orosz@GergelyOrosz·
MCPs are the opposite of dead. They are the life blood of how AI agents use services inside mid-sized and above companies. Case in point: Uber runs on MCPs internally, for good reason. Details: newsletter.pragmaticengineer.com/p/how-uber-use…
@levelsio@levelsio

Thank god MCP is dead Just as useless of an idea as LLMs.txt was It's all dumb abstractions that AI doesn't need because AI's are as smart as humans so they can just use what was already there which is APIs

English
91
100
1.3K
263.9K
Mike Ritchie
Mike Ritchie@thisritchie·
MCP is dead the same way dropbox was in 2007
Mike Ritchie tweet media
Roey Ben Chaim@roeybc

@fatih You can just use the api with info from the swagger or wrap it with a cli. mcp means that every service provider has to add abd maintain another layer on top of that with debatable benefits (context rot)

English
1
0
0
446
Mike Ritchie
Mike Ritchie@thisritchie·
@roeybc @fatih Tell your CFO to "just use the api with info from the swagger". I'll wait.
English
1
0
5
374
Roey Ben Chaim
Roey Ben Chaim@roeybc·
@fatih You can just use the api with info from the swagger or wrap it with a cli. mcp means that every service provider has to add abd maintain another layer on top of that with debatable benefits (context rot)
English
1
0
9
5.6K
Fatih Arslan
Fatih Arslan@fatih·
They are not dead. They are a must have for any enterprise installation and companies with hundreds of employees. I also thought MCP was dead, but you're thinking like an indie hacker. Assuming you're working at a company with 50 people, people install remote MCP servers, use Oauth and then are done. Everyone is authenticated and you don't have to fiddle around with CLI's. It's also secure. For example check @cursor_ai's new Plugin MarketPlace and you got the idea why MCP's are blooming and why they solve a real issue.
@levelsio@levelsio

Thank god MCP is dead Just as useless of an idea as LLMs.txt was It's all dumb abstractions that AI doesn't need because AI's are as smart as humans so they can just use what was already there which is APIs

English
46
27
431
116K
Mike Ritchie
Mike Ritchie@thisritchie·
@karpathy @nummanali also saw your note about SSH. one thing I like about this is "presets" which start a terminal in a predefined dir with preset bash
Mike Ritchie tweet media
English
0
0
0
18
Mike Ritchie
Mike Ritchie@thisritchie·
@karpathy @nummanali built my own (Butter) with tauri. toggle / hide controlled with tags in sidebar, but you can close sessions and reopen them with saved state
English
1
0
2
742
Numman Ali
Numman Ali@nummanali·
Claude Code teams with tmux is really cool When you run with team mode enabled in tmux, it automatically opens the additional terminal in pane I don't really get my main agent to orchestrate, I chat to them myself CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=true claude
Numman Ali tweet media
English
62
75
1.4K
187.1K
Todd Saunders
Todd Saunders@toddsaunders·
@therick Why is CC a nonstarter? This sounds like a bug, not a feature
English
3
0
11
10.8K
Todd Saunders
Todd Saunders@toddsaunders·
No shame on Replit, built a bigger business than I've ever built. But at the rate Claude Code /Cowork is moving, why does Replit needs to exist? What's the bull case for the future of Replit and similar tools if all the model companies are on the same path.
English
187
8
618
175.5K
Mike Ritchie
Mike Ritchie@thisritchie·
@samuelcolvin @stripe I've always been a little unclear on the use case for Sigma. It seems if you have enough stripe data to warrant using it, you'd already be syncing the data to a warehouse / data lake.
English
0
0
0
417
Samuel Colvin
Samuel Colvin@samuelcolvin·
I've long been a fan of @stripe but their Sigma AI is one of the most disappointing AI experiences I've seen recently. The UI is poor, but the real problem is it's not agentic! I ask it > show me the number of customers charged on the different tiers (team and growth) in each month this year Instead of going and finding the plans and inferring what I mean, it can only try to one shot the solution, and of course fails. Come on Stripe! @patrickc can I show you @pydantic AI?
Samuel Colvin tweet media
English
3
3
43
31.7K
Mike Ritchie
Mike Ritchie@thisritchie·
@thenanyu when you're iterating on a problem, you might leave artifacts of a dead end. 1. start down path A 2. switch to path B 3. some stuff from path A is used in B, some isn't i don't want the agent to automatically nuke path A until I've committed to path B
English
0
0
0
110
Nan Yu
Nan Yu@thenanyu·
I see things like /simplify and the existence of code review and bug finding AIs. I have to ask, why do these things exist? Why doesn't the coding agent just naturally do these things? I'm sure there's a good answer. Can someone help me understand?
Boris Cherny@bcherny

In the next version of Claude Code.. We're introducing two new Skills: /simplify and /batch. I have been using both daily, and am excited to share them with everyone. Combined, these kills automate much of the work it used to take to (1) shepherd a pull request to production and (2) perform straightforward, parallelizable code migrations.

English
139
5
518
155.8K
Henrick Johansson
Henrick Johansson@compliantvc·
@sama This sounds like GDPR violation galore. We'll see what the EU has to say about this.
English
88
13
1K
199.2K
Sam Altman
Sam Altman@sama·
Tonight, we reached an agreement with the Department of War to deploy our models in their classified network. In all of our interactions, the DoW displayed a deep respect for safety and a desire to partner to achieve the best possible outcome. AI safety and wide distribution of benefits are the core of our mission. Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement. We also will build technical safeguards to ensure our models behave as they should, which the DoW also wanted. We will deploy FDEs to help with our models and to ensure their safety, we will deploy on cloud networks only. We are asking the DoW to offer these same terms to all AI companies, which in our opinion we think everyone should be willing to accept. We have expressed our strong desire to see things de-escalate away from legal and governmental actions and towards reasonable agreements. We remain committed to serve all of humanity as best we can. The world is a complicated, messy, and sometimes dangerous place.
English
15.9K
4K
34.2K
38.1M
Mike Ritchie
Mike Ritchie@thisritchie·
@BenyaminHolley > Once you start scaling you'll see how their reporting breaks their API gives you everything you see in the app. load it all to duckdb or @definiteapp and go nuts.
English
0
0
0
28
🏍benyamin
🏍benyamin@BenyaminHolley·
lots of people in the replies clowning on me for recommending salesforce. they recommended attio etc etc... Attio is a great tool. it is not for anything above 20 people. Once you start scaling you'll see how their reporting breaks, how their automations no longer work and all the other things. Salesforce is just a bunch of apex code in a trench coat. It's not a real "SaaS" platform even though it was the first SaaS platform. It's 100% possible to manipulate it totally headless. So, for everyone saying "Salesforce? LOL" please get to 10m in rev and get back to me. I'm not saying that in a dismissive way, I'm just saying that Attio literally doesn't scale.
🏍benyamin@BenyaminHolley

A friend at a nine-person startup just asked me what tools he needs to build out his GTM stack. He rattled off RB2B, Clay, Apollo, Instantly, HeyReach, Nooks. My answer was basically: slow down. If you're a nine-person company, you need a CRM. That's your foundation. Everything else is optional until you've proven your sales motion actually works. I recommend Salesforce. I've built Salesforce orgs from scratch and HubSpot orgs from scratch, and I have a strong preference. DM me or argue in the comments. You can hook up the Salesforce CLI to Claude Code and brain dump about how you want your CRM to work. Deal stages, fields, object relationships. It builds it. You can upload CSVs, have Leadmagic enrich them, and push data back into Salesforce through the CLI and Apex code. Plain English in, production-grade CRM infrastructure out. You can't do that with HubSpot. They don't have that programmatic layer. People complain about Salesforce being clunky, but in the AI era, that underlying architecture is actually an advantage. The complexity that made it annoying to configure manually makes it incredibly powerful when you can just talk to it. Sorry Hubspot fanbois ¯\_(ツ)_/¯ Stack for a small team: Salesforce, Leadmagic, Claude Code. That gets you surprisingly far. You could use just that to get to 1m ARR tbh RB2B and website de-anonymization tools are only useful if you already have meaningful traffic. Even then, you need to layer multiple providers, same as an email waterfall. The problem is most of these providers license data from the same handful of upstream sources. You're paying five vendors to validate the same email from the same dataset. It's basically a Ponzi scheme for email data providers. I'm being hyperbolic, but if you don't understand how these systems work under the hood, you're subsidizing everyone who does. LinkedIn automation: I use Lemlist. The specific tool doesn't matter. What matters is you cannot treat LinkedIn like an email campaign. LinkedIn monitors everything. I send maybe 5-6 highly relevant messages a day, only to people I'm fairly confident will respond. Your LinkedIn profile is one of the most valuable things you own professionally. Don't blow it up trying to scale before you know what works. On dialers: don't. If you're a ten-person company, have people dial off their cell phones. Why does a founding AE need recorded calls? Who's reviewing them? Tools like Nooks and Gong exist so managers can verify reps are actually working. They're built for the professional managerial class, not for the person actually selling. Don't go buy Nooks and all this other shit. Most of us would be better off with a notepad, a pen, and a cell phone. Get your CRM right. Hook it up to Claude Code. Add a basic enrichment layer. Go talk to people. Don't make sales rocket surgery before you even close any deals. That's just MBA level procrastination.

English
13
0
24
4.5K
Mike Ritchie
Mike Ritchie@thisritchie·
@karpathy i think the small percent left that it still consistently gets "wrong" is mostly a matter of context and personal preference.
English
0
0
0
299
Andrej Karpathy
Andrej Karpathy@karpathy·
It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the "progress as usual" way, but specifically this last December. There are a number of asterisks but imo coding agents basically didn’t work before December and basically work since - the models have significantly higher quality, long-term coherence and tenacity and they can power through large and long tasks, well past enough that it is extremely disruptive to the default programming workflow. Just to give an example, over the weekend I was building a local video analysis dashboard for the cameras of my home so I wrote: “Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me”. The agent went off for ~30 minutes, ran into multiple issues, researched solutions online, resolved them one by one, wrote the code, tested it, debugged it, set up the services, and came back with the report and it was just done. I didn’t touch anything. All of this could easily have been a weekend project just 3 months ago but today it’s something you kick off and forget about for 30 minutes. As a result, programming is becoming unrecognizable. You’re not typing computer code into an editor like the way things were since computers were invented, that era is over. You're spinning up AI agents, giving them tasks *in English* and managing and reviewing their work in parallel. The biggest prize is in figuring out how you can keep ascending the layers of abstraction to set up long-running orchestrator Claws with all of the right tools, memory and instructions that productively manage multiple parallel Code instances for you. The leverage achievable via top tier "agentic engineering" feels very high right now. It’s not perfect, it needs high-level direction, judgement, taste, oversight, iteration and hints and ideas. It works a lot better in some scenarios than others (e.g. especially for tasks that are well-specified and where you can verify/test functionality). The key is to build intuition to decompose the task just right to hand off the parts that work and help out around the edges. But imo, this is nowhere near "business as usual" time in software.
English
1.6K
4.8K
37.3K
5M
Mike Ritchie
Mike Ritchie@thisritchie·
@beefan they will always find one more thing to add as CRITICAL
English
1
0
0
62
Ethan Cohen
Ethan Cohen@beefan·
tbh i'm very close to being done with AI code review. i KNOW that it's just LLM non-determinism but: - the infinite loop of iteration is so much worse than before - false positives make noise:signal harder and slower to parse - the provider feedback loops seem useless
English
2
0
2
192
Samuel Colvin
Samuel Colvin@samuelcolvin·
I've connected claude code to @gmail (via a simple CLI) and it's fucking magic. "Who's our account manager at XXX?" "Find me the email of every VC who's reached out in the last three months, write them to a CSV file."
English
17
5
113
18K
Mike Ritchie
Mike Ritchie@thisritchie·
@wesmckinn that's my go to example for software not being "solved". nailing all the details for end user experience is still very much a human job.
English
0
0
0
115
Wes McKinney
Wes McKinney@wesmckinn·
Claude Code for web is a hot mess. I tried it today and it got stuck forever with some GitHub App error. Not ready for serious use
English
8
1
19
5.6K
Mike Ritchie
Mike Ritchie@thisritchie·
> Lead databases — Clay, Apollo, RB2B will be fine. Claude can't collect 1B phone numbers and emails. telling Claude to run web searches is less friction and often better than the data in Apollo. e.g. if I want # of employees, current job openings and amount raised for a list of companies.
English
0
0
1
487
Mike Ritchie
Mike Ritchie@thisritchie·
Great example of how fast AI is moving and how hard it is to grok all the moving pieces. > MCP is dead in the water Obviously false, millions of users of MCP protocol every day and growing fast (even if the user doesn't know they're using MCP) > every MCP server you connect loads its tool definitions into your context window. name, description, parameter schema, all of it. connect 10 servers with 5 tools each and you've burned 50 tool definitions worth of tokens before your conversation even starts. This was true in 2025, but mostly solved in 2026. Claude Code MCP Tool Search launched on January 14, 2026, which completely solved the problem in Claude Code. Other products will adopt the same pattern. To pile on, 1M token context windows will become the norm this year (already is in some models), so this argument is bogus. > i assume this is why @steipete left it out of @openclaw OpenClaw barely existed in 2025 (Jan 2026 was the takeoff) and now many are assuming the decisions made on OpenClaw should be applied universally (they should not). I'm not even terribly pro-MCP, but this tweet has 1k likes and is completely wrong. The funniest part is it was sort of right in November 2025, but OpenClaw didn't exist then.
Chrys Bader@chrysb

unpopular (maybe?) opinion: MCP is dead in the water @openclaw has shown me that api & cli will win. every MCP server you connect loads its tool definitions into your context window. name, description, parameter schema, all of it. connect 10 servers with 5 tools each and you've burned 50 tool definitions worth of tokens before your conversation even starts. context bloat will never be a good thing - performance-wise or economically. i assume this is why @steipete left it out of @openclaw. the "exec" tool paired with on-demand skills is all you need. it can run any command invented since the beginning of computers. a resurgence of glory for ancient, but powerful tools like curl, sed, awk, grep. command line tools once mastered by the greats, but long forgotten and buried underneath abstractions developed for us lesser mortals. now available to us all, piloted by the smartest models on earth. every founder gets their own mass army of greybeards. the inertia required for MCP adoption, imo, is too great to overcome the momentum @openclaw has breathed into api + cli + skills. the common defenses people bring up: • "MCP gives you typed schemas and validation" — so does a well-documented CLI • "MCP gives you explicit permissions" — so does a sandbox with an allowlist • "MCP is a standard" — a standard that scales poorly is still a standard that scales poorly lastly, i've heard many MCP servers are just wrapping existing APIs - that kind of redundancy and unnecessary indirection should be a red flag. so, let's drop it and redirect our efforts into cli tools & apis with accompanying skills.

English
2
0
0
319
Mike Ritchie
Mike Ritchie@thisritchie·
@chrysb @openclaw is this engagement bait? looks like it's working. mcp context window bloat is a solved problem.
English
1
0
2
112
Chrys Bader
Chrys Bader@chrysb·
unpopular (maybe?) opinion: MCP is dead in the water @openclaw has shown me that api & cli will win. every MCP server you connect loads its tool definitions into your context window. name, description, parameter schema, all of it. connect 10 servers with 5 tools each and you've burned 50 tool definitions worth of tokens before your conversation even starts. context bloat will never be a good thing - performance-wise or economically. i assume this is why @steipete left it out of @openclaw. the "exec" tool paired with on-demand skills is all you need. it can run any command invented since the beginning of computers. a resurgence of glory for ancient, but powerful tools like curl, sed, awk, grep. command line tools once mastered by the greats, but long forgotten and buried underneath abstractions developed for us lesser mortals. now available to us all, piloted by the smartest models on earth. every founder gets their own mass army of greybeards. the inertia required for MCP adoption, imo, is too great to overcome the momentum @openclaw has breathed into api + cli + skills. the common defenses people bring up: • "MCP gives you typed schemas and validation" — so does a well-documented CLI • "MCP gives you explicit permissions" — so does a sandbox with an allowlist • "MCP is a standard" — a standard that scales poorly is still a standard that scales poorly lastly, i've heard many MCP servers are just wrapping existing APIs - that kind of redundancy and unnecessary indirection should be a red flag. so, let's drop it and redirect our efforts into cli tools & apis with accompanying skills.
Chrys Bader tweet media
English
283
89
1.6K
330.6K
Mike Ritchie
Mike Ritchie@thisritchie·
I am "listened to the radio to find out if school was canceled" years old
English
0
0
1
66