Redis devrel stealing shock content to juice engagement on LinkedIn. Who is approving this. This isn’t it. We have got to do better.
My infra guy explaining an AWS feature to me: “It’s like a GROUP BY in SQL” instantly clicked, thanks for speaking my language
New logo. Hand made, pasture-raised, grass fed. Welcome to Downlink.
The latency and cost savings of small models is substantial. 100x faster and cheaper. The issue is dev time. Time spent curating, training, evaluating, and deployment. What if the small models trained automatically? I’m building this.
10x more politics posts on my Bsky feed than on Twitter. The custom feeds don’t help.
We need a social credit score for restaurants. I don’t customize or ask questions, I just point at the thing I want. I deserve some recognition. I should get to cut the line.
Silo season 2 It was 4 episodes too long, way too dark, and completely threw out the idea that it takes DAYS to walk up. The ending was cool at least.
One more time for the people in the back: YOU CANNOT ASK LLMs FOR A CONFIDENCE SCORE. THEY DON’T KNOW WHAT THAT MEANS.
Scott McGrath@smcgrath.phd · 2y ago🧪 LLMs ace medical exams yet fail to recognize uncertainty, risking patient safety in clinical settings. Using MetaMedQA, researchers exposed a gap between AI confidence and accuracy, urging improved metacognition for safe healthcare deployment. 🩺💻 #MLSky
VCs making fun of first responders in the middle of a major emergency makes me sick. What the hell is your problem. Also, no one gives a fuck who your daddy was.
Omg omg omg omg omg!!!
Martin Kleppmann@martin.kleppmann.com · 2y agoGot a bottle of wild boar wine for Christmas! Does the picture on the label seem familiar? In case you’re wondering, it’s taken from an out-of-copyright book: General Zoology by George Shaw, vol. ii, part 2, page 458, published in 1801. archive.org/details/p2ge...
I was NOT prepared for when nostalgia came for me. The resurgence of early 2000s music is firing long-dormant neurons in my brain.
When I read technical books I drink wine and take notes on index cards
"Data is the fossil fuel of AI" Please don't encourage the VCs. This phrase will launch a thousand bad blog posts.
Sung Kim@sungkim.bsky.social · 2y agoPre-training as we know it will end - Dr. Ilya Sutskever at NeurIPS 2024
Titan is working on something new 👀 If you make a lot of API calls to gpt-4 or claude and struggle with latency, I’d love to chat. For discriminative-style tasks I see a 30% effective speed up on average. Sometimes much higher.
Titan is working on something new 👀 If you make a lot of API calls to gpt-4 or claude and struggle with latency, I’d love to chat. For discriminative-style tasks I see a 30% effective speed up on average. Sometimes much higher.
This is my preferred way of doing things right now. The CLI only selects the config. All of your config params are grouped together in a single object. You get nice type hints and validation from pydantic. Your script contains a record of previously-run configs.
![Screenshot of the following code:
import click
from pydantic import BaseModel
class Config(BaseModel):
table: str
start_date: str
CONFIGS = {
"payments": Config(table="payments", start_date="2021-01-01"),
"refunds": Config(table="refunds", start_date="2022-02-01"),
}
@click.command()
@click.option(
"--config",
"config_name",
required=True,
type=click.Choice(list(CONFIGS.keys()), case_sensitive=False),
)
def main(config_name: str):
config = CONFIGS[config_name]
print(f"Querying table {config.table} from {config.start_date}")
if __name__ == "__main__":
main()](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:pv6dad2uw2cit2fffq5jly34/bafkreieumhy3joa4cwniogtdmsu7pjgffnrzfjupbw7ykdxnqjubw6j7ci)
Ethan@ethanrosenthal.com · 2y agoStop putting CLIs on your scripts. CLIs are not reproducible. You’re never going to remember the 6 arguments you fed into the script. CLIs are for libraries, not for running a job.
Can someone check in with Gartner? They’re not okay. They are crying out for help.
The Discover feed here has been complete garbage for me. Would love to hear your custom feed recommendations! Looking for technical content.
If you do a lot of admin/automation with Redshift, these system views are really handy, check it out:
Ladies and gentlemen – I am the median. My “security posture” is within 51.07 percentile. I want to whoever thought this email required such precision.
I’ve seen many good data people and engineers fall into the trap of thinking being right is enough, but being effective is the real unlock to career growth. Wrote some thoughts here: open.substack.com/pub/pedram/p...
It's not enough to be right
Early in my career, I was rewarded for being right.
open.substack.com
When we talk about "references to a research paper" we really need to include all those obscure tech blogs written by 10x engineer anons that you see on Hacker News Google Scholar really should serve me that post from segfault_sommelier00 debunking the paper's claims