Thomas - sourcetms

if you're a vc and you're on copilot you should close your fund

Yesterday I had a further glimpse into the crossover of IT and and business in major enterprise (5-10k+ employees). Without a doubt every single one of them is leaking sensitive data by the minute due to shadow spend via OpenAI/Anthropic.

Not a worry in the world from employees. Industry agnostic. It makes SO much sense to build an internal AI platform internally vs use a vendor.

sells results as a service better for positioning in software market, larger labor market comparative to software spend businesses and ppl buy outcomes anyway. why not you? why get a seat on a saas?

Keeping up with all the news coming out is difficult. So I spent last night tinkering around and built a 24/7 fully automated news site for tech with an AI focus. Pretty happy with how it came out. Now that the foundation is done, the fun bits are possible 😄

distribute.softwareDistribute SoftwareThe main technology stories of every day, on one calendar: Hacker News, Reddit, the technology desks of major newsrooms, and releases from the AI labs themselves.

it's over at 🔗 distribute.software/ (i had to use the domain for SOMETHING) ...leaving out X and research for now but am sure there's a place for it soon. (mostly on cloudflare w/ the new deepseek search API)

interesting just noticed this in the cheeky pint episode > zuck doesn't use an iphone. i wonder why.

Zuck: "...future where, basically everybody is going to be writing tons of code, they're just not going to know that they're writing code" yes. prompting an agent, building a workflow, tweaking an automation. it's all code, people just don't call it that.

Coinbase fine tuned an open-weights model (Qwen3.5-9B from Alibaba) on its own internal data, and it beat SOTA frontier models on fraud prevention. Being both cheaper and faster too. 🤑 This will no doubt be the continuing trend within companies who take AI seriously.

The private, cloud served, frontier AI models are amazing at broad tasks and reasoning, but they obviously can't compete on internal, domain-specific narrow tasks. There's little comparison to a specialist model.

Those are the models that are increasingly winning inside large co's. 12-18 months from now we could have fine-tuning done by technical employees at (non frontier) companies 👀. Let's see.

Making content on every other platform feels so normal after you get into it, but LinkedIn still feels like this self-serving, circlejerking, awkwardly labelled, professional, self inflating environment that just makes no sense. And it is so cringe to post there. I don't get it.

2 weeks ago we took the material we teach inside companies and put it online for free. how to build an AI strategy. how to roll out AI technically and operationally. how to design an internal FDE team. how AI changes your company.

> how do we run a champions program so non-technical people build their own workflows > who in the company owns this in a few companies people are literally fighting over who owns AI. resources will be linked below.

specialtytokens.com/resources

specialtytokens.comSpecialty Tokens - AI Training and Implementation for Enterprise Teams - US, UK, EU - EN/DEA senior team across strategy, integration engineering, agents, and rollout. Advisory and hands-on delivery that makes enterprise teams AI-native.

it used to live in private workshops. now it's a public library of everything we've learned. 600 people from all over the world grabbed it in 2 weeks. screenshots below are from our internal dashboard. people want to know: > how do we build an AI strategy

remember when openAI and everyone went all studio ghibli?

The best "secret" of ChatGPT/Claude? Programmatic or vision based computer use. Too lazy to do something? Fill a form? Dear mr computer pls do it. Watch the computer us the computer

Pretty impressive to see how tools can make us more effective in our work. //Who would have thought? The average accountant without AI assistance scores around 37% on core tasks. Using AI (Opus 5 in the top right) shows an incredible leap in performance on tasks.

I believe that most AI tools with the correct human guidance, creativity, and technical environment can help us handle 20x the volume.

To be honest, I think this applies for almost all knowledge work.

* I would still want a real accountant to be responsible and licensed. At the end of the day it makes sense to engage with a person who knows the system and the tasks and WHY and just not a system.

companies are planning AI for Q4/Q1 right now. here's the maturity patterns I keep seeing on calls, and what actually moves a company up a level:

most companies I talk to sit at 1 or 2. almost none are one level company-wide. marketing is a 4 while finance is a 0.

the leap that makes a difference is step 2 to 3 usually. it's one person whose responsibility and job is to help everyone else build and standardize.

companies are planning AI for Q4/Q1 right now. here's the maturity patterns I keep seeing on calls, and what actually moves a company up a level:

Level 4: one governed source of company info for agents and chat. teams build their own workflows and agents, then hand them to IT to professionalize. tool sprawl still exists, but usage keeps climbing.

Level 5: the company ships agentic products externally, with a dedicated AI lead and internal forward deployed engineers. might experiment with models and fine tuning > more advanced AI builds.

companies are planning AI for Q4/Q1 right now. here's the maturity patterns I keep seeing on calls, and what actually moves a company up a level:

Level 2: lots of subscriptions, vertical tools per department, consistent training, light integrations.

Level 3: someone actively owns AI. a volunteer applied AI team or an official unit. skills and connectors work and the central team helps everyone build.

companies are planning AI for Q4/Q1 right now. here's the maturity patterns I keep seeing on calls, and what actually moves a company up a level:

Level 0: no subscriptions bought. employees use personal credit cards and company data leaks. this is a procurement failure, not an employee one.

Level 1: company bought a plan. everyone copy-pastes between apps, MCPs/connectors sit unused. maybe a weekly AI learning hour.

xAI, OpenAI, and Anthropic all pushed for slowdown rhetoric. then immediately competed on who could ship the best performance per dollar. the ask was never about slowing down their own work. it was about regulating everyone else. classic.

22 Sept

ai product scanners are replacing consultants. they audit your entire stack automatically. posthog has self-driving analytics. security scanners catch vulnerabilities. code quality scanners flag tech debt. design system scanners check consistency. the consultant showed up once and left.

22 Sept

the scanner runs every single day.

20 Sept

Jev truly is a sign of Jevon's paradox. People want to optimize and continue to use AI everywhere they can.

I started using Jev as a tool call for our users on supercenter. And oh boy are they gonna love this.

supercenter.appSupercenterBuild, run, share, and govern AI and agents across your company, with central control over access, spend, and data.

Agentify your operations with the tools you already use. see: supercenter.app

your agent tools are going to get way way faster. thanks to jev. the #1 use case in enterprise software is classification.

two days ago a former OpenAI researcher (worked on ChatGPT's RLHF) released jev. it's a model that can't talk back. here's the problem it solves:

LLMs write and explain. jev decides and scores. use ChatGPT to draft your email. use jev to route 10,000 tickets while you sleep. their numbers: 194x faster, 445x cheaper than LLM baselines on classification workflows.

$0.042 per million tokens. 70-500ms latency. output tokens are free because it doesn't write prose. if judgment becomes cheap enough to run on every ticket, every agent step, every tool call... the amount of AI in software (again) goes up, not down. that's the jevons paradox bet.

two days ago a former OpenAI researcher (worked on ChatGPT's RLHF) released jev. it's a model that can't talk back. here's the problem it solves:

every time you ask ChatGPT to categorize something or make a decision, you get back a paragraph. then you have to read it, interpret it, and hope it's in a format you can actually use. jev doesn't write paragraphs. it just answers questions with structured data.

you give it a customer support ticket and ask: is this urgent? what category? what's the sentiment? it returns: urgent (87% confident), billing issue, frustrated. there's no explanation. no longwinding essay. just decisions your systems can act on immediately. the difference:

14 Sept

I think this sums up those slowdown posts from Dario/Elon/Sam.

10 Sept

i just realized when people say software factory, agent deployment center, control plane, and ai gateway they literally all mean the same thing.