Powerful GPUs. No infrastructure setup. Run models through tools like Ollama for inference during development (eg. synthetic data generation).
Notebooks are .py files. Git-native. Import functions between them. Share experiments as live apps with one click.
Would you be up for a quick 15-min demo?
Could I fully replace `ollama` with `llm` ?
Enverge Lab = GPU + next gen notebooks + git + demo experiments
No, sorry. I'm not releasing an Excalidraw #course soon.
LLMs don't build better software, they build bad software faster.
As tools, LLMs discourage understanding in a way that no other tool does.
The indicators are not the kids using chatGPT for dating advice.
Or in product managers vibe-coding prototypes.
They are in the hard upsell Google does with Gemini. Or in the random "AI" text on unrelated products.
vibekanban.com
So, given the amount of code + features that will be generated by this calss of tools, the market broadens. If fans of Adam Smith economics are right (and they have, for the most part so far), specialised engineers should become more expensive in a few years.
A win is a win, right?
"We're a startup focussed (sic) on AI and programming languages, based in ... backed by leading investors including Y-Combinator."
red flag 🚩
odds are they will be gone in 6 months
The only benckmark AI tools should be track is solving gigs on upwork.com .
Co-working spaces when most people are on calls gives me call center vibes.
Am alone on this?
Substack seems to be a better platform for long form ideas, so I decided to republish my "Beyond the Hype: Your Complete Guide to ML Development" article here.
Too bad tables can't be embedded.
open.substack.com/pub/enverge/...
Recently discovered Kdenlive. Such a beautiful and useful open-source tool, especially for founders who record a lot of video tutorials.
Would recommend. #founders
... camera in their pockets didn't make cinema obsolete, it only pushed creativity to new levels and widened the spectrum.
AI can 10x the 10x engineer.
Regardless of improvements, LLMs alone are loose cannons.
A powerful tool is still a tool and is wielded by the craftsman.
🌶️ #AI #Vibe
"I can't wait for AI to replace all these naggy engineers" is how most posts and attitudes come across.
My popcorn is ready for the moment when vibe coders wrap their heads around concepts of higher abstraction layers, systems design trade-offs, and graph theory.
Everyone having a ... 👇
Nothing makes a founder eat humble pie as the moment they realise poor tech decisions.
A rewrite is sobering.
don't tell anyone, but Dash + Cursor is a whole new level of hack for building web apps.
what are you using on the front-end?
The Jupyter ecosystem has quite a broken multi-tenancy model.
BRB, fixing it
Today Claude said the ultimate developer lie: "Let me make one final modification". It has been trained well.
My biggest takeaway? Validating and invalidating assumptions - seeing which ones held up and which ones crumbled.
Time to build something better.
Mental health matters as much as the pressure you face.
This week, I stepped away from daily ops. I volunteered at QCon, helping with the small details that make conferences special. Connected with attendees, speakers, and sponsors.
Some basic logic for why OpenAI will probably release an ~30B param reasoning model with MIT/apache.
* OpenAI will only release something clearly SOTA in size category
* Will release a reasoning model, else hard to tell story around popular evals
* Want people to use it, cant be too big
...
2 hours. 17 cloud resources. One region to another.
All because we ran out of GPUs.
No dropdown menu. No simple fix. Just config updates + trial and error.
This is why we're building Enverge.ai. Zero cloud config. Zero headaches.
Just training.
#CloudComputing #AI
lorem ipsum was a vibe
On the rare occasion I needed gen AI to output images with bad text, it actually fails to do so. The text isn't unreadable gibberish.
You have a brilliant idea for fine-tuning an LLM model.
What tool do you reach for first?
1. Google Colab
2. A self-hosted Jupyter instance
3. an IDE
Would love to hear your preference, because I've been oscillating between a few options recently.
Is this plain full-text search or is it powered by some LLM?
which one are you referring to?
Fine tuning my "fake account detection" skills by browsing LinkedIn. Wild.
what infra provider are you using for these 8 H200s?
