
FDE at @collide_ai. I build production AI for document-heavy work. 20 years in oil & gas, SEO, & SaaS. 7th gen Texan, family ranch steward.
East Texas
Joined March 2026
- Tweets283
- Following677
- Followers208
- Likes758
Pinned Tweet
Everyone wants to deploy AI agents in oil and gas.
Almost nobody does the boring part first.
12,000 land and title docs. Scanned, handwritten, never linked.
We turned them into one graph:
- every well tied to its lease
- every lease tied to its tracts
- every answer traced back to the page it came from
The trick is you solve the data problem before the agent ever touches it.
Agents are easy. Trust is the hard part. teapawt.net/FracSlap/status/210611…
Murdoch Murchison retweeted
We had claude and codex rewrite our app in rust. No tests anymore. If it compiles, then we ship.
And no PRs. Just merge onto main.
Can you imagine how insufferable it would be to work with an actual rust developer?
Thankfully claude and codex make far fewer mistakes than most devs.
Get on the short bus 🚍
Even though this deck doesn’t have a screenshot or a single mention of the landman Gopher (typical), I still endorse it. Riggs is tighter than @FracSlap’s fade on cut day.
Y’all don’t forget to take your ai pills this morning.
At @collide_ai we’ve graduated to mobile iv bags of ai juice, mainlining 24/7.
$360,000 a year to host our model on a cluster of A100s
We sharded the model to four L40 gpus and cut the cost to $56k
Then we extracted the knowledge out of the weights, baked it into the harness, now we can fit a smaller model on one gpu to have it fun domain specific tasks.
Collide team is cooking
Last week I ran a mineral interest back to patent (1892), then shipped the code that does it for the next ten thousand.
Landman by trade. Builder by nature.
Almost like the two paths collided.
Uncle Chuck’s parallels are second to none. Read 👴☕️
Call it whatever you want.
Anyone can learn to prompt Claude in a week.
What you can’t do is know which mistake in the customer's corpus is the expensive one. Like a royalty decimal that's off by one digit.
20 years in oil and gas taught me that.
Not Claude.
Forward deployed engineer is a stupid job title!
Are we in the military deploying the troops to build AI agents?
You’re a solutions architect who knows how to prompt Claude. Don’t let the propaganda from Palantir get to your head.
Some of you aren’t even solutions architects. You’re consultants pretending to be more important than you actually are!
Can we please go back to job titles that actually describe the work that is being done and not some weird pseudo-militaristic hype that we are all swallowing right now?
A bank putting AI on knowledge work still needs every answer boxed to the exact words it came from.
The partnership gets you in the door.
The person who knows when a number is wrong is what keeps it in production. teapawt.net/gdb/status/21184108189…
Announcing our strategic partnership with JPMorganChase, to explore how advanced AI can support their enterprise: helping engineers build, security teams defend, and employees and customers benefit from more capable AI.
I'm especially excited about working together on cybersecurity, to help secure and protect the financial system to a level that is only possible with AI.
65
Five models pulling lease terms on their own: as low as 3 to 42%.
Same models, with a field contract written by land experts: 86 to 88%.
The fix wasn't a bigger model. It was people who've read deeds writing the instructions.
StrataBench is free. Try it on your own docs. teapawt.net/FracSlap/status/211803…
The oil and gas industry is built on land and title documents but AI models struggle to answer questions that landmen ask every day.
Today we're releasing our land and title benchmark to the public for anyone to use for free.
Stratabench is an evaluation suite built to score any AI model on how well they classify, extract, and define data from land and title documents.
It's free for anyone to use on huggingface (link in the comments)
1
57
In Anthropic's agent study, 77% of analysts reused the AI's argument in at least one answer. Only 15% credited the AI anywhere.
You just stop noticing whose idea it was.
Tie every answer to its source.
If the system doesn't keep the receipts, nobody will.
2
54
Source: anthropic.com/research/agent…
29
Murdoch Murchison retweeted
Most well intervention is reactive. Blue Spark wants to know what's coming: shatter scale with electric pulses, then bring samples to surface. CTO Matt Mohajer said they aimed to launch it at ADIPEC, happening this week. EP 87: youtu.be/7SDJ1OzpKdQ
Anthropic gave 50 expert analysts an AI agent with coin-flip opinions. They sided with the coin 86% of the time.
"Human in the loop" isn't a plan. A reviewer who agrees with everything is a slower Approve button.
Real review needs someone who knows the work enough to say no.
1
1
92
Most of these pullbacks aren't AI failing. They're AI shipped with nobody checking its work.
In O&G a wrong number on the wrong well doesn't look like an error until it's expensive. The fix isn't more seats. It's someone who knows the business sitting next to it. teapawt.net/unusual_whales/status/…
BREAKING: Employers worldwide are pulling back on AI as they navigate a flood of errors and low-quality work that sap employee time, raise privacy and security risks, and drive up costs, per the Guardian.
1
59
29 agents on one 16-CPU box while Collin eats dinner.
The number I like is 3: merges the swarm refused and sent back. Zero halts.
Speed is easy now. A reviewer that's willing to say no is what makes it something you can actually trust. teapawt.net/FracSlap/status/211564…
29 agents working while I'm at dinner.
6.5 hours in, 8 hours to go.
Running into compute constraints on testing, swarm is hosted on a 16cpu 128GB machine.
we are only limited by our creativity and compute--welcome to the future.
2
4
501
Murdoch Murchison retweeted
Average American business still prints the PDF to sign it. Ask Midland.
OpenAI's James Sun and Ari Weinstein on the AI adoption gap between San Francisco and the average American business, and why agentic computer use is the way to bridge it:
"Living in San Francisco, we see a lot of thinking, 'Wow, the labs can do everything,' and, 'Why doesn't everybody just install Codex and just have it do stuff?'"
"But for the average American and the average small to medium business owner or the average enterprise, even understanding what the tools are capable is [hurdle] one, connecting tools in a way that makes sense is two, deploying it is three, servicing is four."
"The really cool thing about computer use as a technology is the universality of it. It works with every application. It's unlike any other thing that we've built where the developer has to do some work to adopt it, someone has to build a connector or MCP server."
"If there are complex workflows that require a lot of disparate software, computer use is a great way to join them, complete the entire end-to-end workflow, and then building it in a format and UI that feels good to hold for those customers feels still like a really, really big opportunity that I think is still open."
@JamesZmSun @AriX
OCR gets you the text. That's the easy part.
In oil and gas the same well shows up in a permit, a division order, and a production report with three different numbers.
The hard part is the AI knowing which one wins. teapawt.net/oliviscusAI/status/211…
You can turn scanned PDFs, tables and equations into text your AI agent can actually use.
It’s called MinerU. It converts PDFs, images and Office documents into structured formats like Markdown and JSON.
- Extract text from scanned pages with OCR
- Handle multi-column layouts
- Convert tables into HTML
- Export recognised equations as LaTeX
- Process documents in batches
- Run locally without uploading files
Use it through the CLI, Python SDK or web interface, then feed the results into your research or document workflows.
Less copying from PDFs. More time working with what’s inside them.
1
4
180
Here's what that looks like when the documents are deeds and probates instead of PDFs of papers:
teapawt.net/MurdochMurch/status/21…
Every chain of title is a scavenger hunt.
A patent. Some deeds. A probate two counties over. A correction deed fixing the last correction deed.
The trick isn't reading the documents. It's putting them in order and proving every link.
1
24



















