As many of you know, I have long been an advocate for Apache Spark and Databricks. Having conducted extensive work and authored two books on the subject in my previous life, it's exciting to see our paths seem to converge once again at the intersection of Data and AI. I believe that Generative AI and LLMs will transform the current data lake into what I like to call "Lake Vectoria," a lake of embeddings.
This blog post shares my initial thoughts on Databricks' latest open LLM offering, DBRX. As I explore this model further, I will continue to share more insights. Stay tuned! #llms #dbrx #databricks @matei_zaharia@databricks@Roost
#166 Flexing Bricks as Open Weights: The DBRX by Databricks https://t.co/UPm6IYJBb2
So, selling tools to bankers instead of doing the banker's work is a mistake AI founders may make?
In BFSI, the winners will be AI that:
- processes claims
- reconciles ledgers
- reviews compliance
- underwrites risk
The real change will be regulatory judgment + proprietary data?
9/ Until we teach AI to forget on purpose, every deployed model remains a ticking time bomb of memorized secrets waiting to be extracted.
The future belongs to models that know when to access data, not absorb it.
https://t.co/fA6pXeXPne
The most dangerous thing about enterprise AI isn't what it doesn't know. It's what it remembers too well.
Every language model faces a fundamental tradeoff: memorize in its weights or retrieve from external sources.
For enterprises, getting this wrong can be catastrophic 🧵
8/ The solution isn't preventing memorization entirely. It's teaching AI to forget on purpose.
The winners will master controlled forgetting:
• Continuous evaluation to find risks
• Selective unlearning to remove them
• Retrieval from secure sources instead of memorization
@pitdesi Does it even work? Once the term gets long enough, the math stops caring and gravity takes over. Stretch it to 300 or 3,000 years, and the payment barely moves. At that point, Euler wins.
@pitdesi Yes, LLMs just can’t help themselves, especially ChatGPT. No matter how many times you say “no em dashes” in the instructions, you still have to remind it to remove them before the final output.
OpenAI Dev Day has cemented the platform shift that has been unfolding over the last few years.
The SDLC as we knew it is dead.
A new one is emerging, built around chat interfaces, drag and drop agent builders, and code that can be edited as easily as a script.
This is the new foundation of how software will be built.
#AgenticAI #OpenAI #ChatGPT
#216 The Beginning of the End for IDEs https://t.co/HCg1DvSwMP via @LinkedIn
@sama Good start. Currently it mostly covers the things I would like to explore but hopefully soon it will cover what is (or should be) on top of my mind today.
Picture 4 components: Information, Intelligence, Agency, Action
They COULD connect 12 different ways
Only 4 connections actually work
Why? Constraints:
•Information never pushes (passive)
•Action never pulls (execute only)
•Intelligence can’t decide (no executive function)
92% of AI agents aren’t agents at all.
They’re missing the one edge that matters: Agency→Action
Without it, you have infinite analysis and zero outcomes.
The Cogentic AI Graph explains what everyone’s missing: 🧵
I'm getting seriously frustrated with #Claude overloading all the time on my $100 Max plan. Maybe it's time to bite the bullet and upgrade to the $200 one for that 20x capacity boost. The thing holding me back is that I use #ChatGPT way more every single day, and it never gives me grief about busy servers. Even #Gemini has stopped complaining lately, though it still randomly forgets the context mid-conversation.
Vibe coding: where developers enthusiastically hand over the keys to agents who will definitely, absolutely, totally not do our jobs better than us. #agenticai#vibecoding#agentdriverdevelopment #212 Humans Will Vibe, Agents Will Code https://t.co/AbxHu3tqgs via @LinkedIn