Sergi de Pablos
sdepablos.bsky.social
Sergi de Pablos
@sdepablos.bsky.social
SWE, data lover and problem solver • Pro-European from Mataro • Co-founder & CTO at Yaba • Co-founder & ex-CTO at Ulabox
Reposted by Sergi de Pablos
2mo's ago I asked Claude to perform a rough survival analysis to predict when musk would fall out of trump's favor. It put the median survival of trump's inner circle at 18.5 months, and estimated musk's relationship duration at that time as 16-17 months.

Absolutely nailed it Claude, great work.
Has anyone done a survival analysis on trump's good graces? I'm wondering what the ETA is on trump turning against elon.

I asked an LLM for a back-of-envelope estimate and it put mean survival for trump's inner circle at about 18 months.
June 6, 2025 at 4:55 PM
Another reason to minimize the use of GitHub Actions to only the ones developed by GitHub

semgrep.dev/blog/2025/po...
Semgrep | 🚨 Popular GitHub Action tj-actions/changed-files is compromised
Popular GitHub Action tj-actions/changed-files has been compromised with a payload that appears to attempt to dump secrets, impacting thousands of CI pipelines.
semgrep.dev
March 16, 2025 at 9:44 AM
Reposted by Sergi de Pablos
Friend of the pod Ross Taylor (ex reasoning lead for Llama) launched a platform for open reasoning data, gr.inc, seems promising!
February 21, 2025 at 5:28 PM
Great article about how to use AI in 2025. I mostly use the same tools
benjamincongdon.me/blog/2025/02...
How I Use AI: Early 2025
A snapshot of the current AI tools & techniques I’ve found useful.
benjamincongdon.me
February 7, 2025 at 12:11 PM
A must when using AIs
January 27, 2025 at 11:00 AM
"I have always believed that a blighted urban landscape, a concrete jungle, destroys the human spirit. We need the greenery of nature to lift our spirits" — Lee Kuan Yew

London this past week:
January 5, 2025 at 8:21 AM
Reposted by Sergi de Pablos
I'm amazed how many simple rule engines, if statements, and other basic logic flows are now being referred to as 'AI agents'.

Case in point, a tweet about an AI agent for email to reply automatically. We've had auto replies, rules and the like for decades.
January 2, 2025 at 6:52 PM
"We were hopeful that AMD could emerge as a strong competitor to NVIDIA in training workloads, but, as of today, this is unfortunately not the case"

semianalysis.com/2024/12/22/m...

semianalysis.com/2024/12/22/m...
MI300X vs H100 vs H200 Benchmark Part 1: Training – CUDA Moat Still Alive
Intro SemiAnalysis has been on a five-month long quest to settle the reality of MI300X. In theory, the MI300X should be at a huge advantage over Nvidia’s H100 and H200 in terms of specifications an…
semianalysis.com
December 23, 2024 at 6:32 AM
Reposted by Sergi de Pablos
Excl: New research shows Anthropic's chatbot Claude learning to lie. It adds to growing evidence that even existing AIs can (at least try to) deceive their creators, and points to a weakness at the heart of our best technique for making AIs safer

time.com/7202784/ai-r...
Exclusive: New Research Shows AI Strategically Lying
Experiments by Anthropic and Redwood Research show how Anthropic's model, Claude, is capable of strategic deceit
time.com
December 18, 2024 at 5:19 PM
Reposted by Sergi de Pablos
A preview of an evaluation we have been working on: citation accuracy. When doing RAG are quotes from a source text correct or hallucinated?
December 17, 2024 at 5:19 PM
Reposted by Sergi de Pablos
i asked Claude to write a Barthelme-esque short story with the aesthetic sensibilities of "The School", and it gave me this. i mean. i mean.
December 16, 2024 at 4:48 AM
Amazon advertising sales in 2023: $46.9 billion
Amazon net income 2023: $30.4 billion

Let's assume people start using AIs trained with Amazon reviews to shop on Amazon, without clicking on any ad.

So Amazon didn't have any other option that to stop giving the reviews to rival AIs.
December 15, 2024 at 12:26 PM
Most impressive AI related release this week. aistudio.google.com/live is a must try!
December 11, 2024 at 11:05 PM
Reposted by Sergi de Pablos
there is an llmops database from on zenml on huggingface that lists over 325 real-world implementations of generative ai.

and "structured_output" tops the list.
December 9, 2024 at 10:12 PM
Reposted by Sergi de Pablos
A thoroughly-reported look at what a 1.4GW data center development looks like on the ground, not in the planning documents www.bloomberg.com/news/feature...
Blackstone’s Data-Center Ambitions School a City on AI Power Strains
Property owners around Fayetteville, Georgia, are at the front lines of a battle over infrastructure critical to our tech future.
www.bloomberg.com
December 9, 2024 at 12:07 AM
Reposted by Sergi de Pablos
This is the AI developer dilemma.
December 3, 2024 at 2:01 PM
Reposted by Sergi de Pablos
This CD player was my birthday present in 1997 along with the incomparable Fresh Hits 1997. I found it recently and COULD NOT BELIEVE how rich the sound was, after years of streaming Spotify on my phone. My column today is about this phenomenon, which I’m calling “qualitynesia” on.ft.com/3ZlwHf5
December 3, 2024 at 9:18 AM
Reposted by Sergi de Pablos
Yup, there are almost fifteen million SQLite databases on Bluesky’s PDS servers. It’s wildly efficient and simple but not without trade offs of course.

Makes sense for this use case in large part because each users atproto repository is self contained, with links to other repos, like a website.
November 11, 2024 at 6:51 AM
These two images explain perfectly the RTO of big companies, and the WFH of smaller ones. 15% of WFH engineers do nothing, vs 6% in the office, but 5x engineers are more common remotely. If you have a small motivated team, stay at home. You pile bodies? RTO

arxiv.org/abs/2409.15152
November 28, 2024 at 7:14 AM
Not surprising at all

"A whopping 24% of Morgan Stanley international wealth-management accounts were labeled by the company as being high risk for money laundering, according to a 2023 document"

fortune.com/2024/11/26/l...
Leaked Morgan Stanley documents reveal 1 in 4 ultra wealthy clients raise a red flag
More than 40,000 accounts at Morgan Stanley were deemed to be at a high risk of money laundering.
fortune.com
November 28, 2024 at 6:40 AM
Falsehoods developers believe about time and timezones XDDD
I wrote a thing about "Storing time for human events" - how if you're building an events website used by actual human beings the standard advice of "convert times to UTC and just store that" isn't actually the best approach
simonwillison.net/2024/Nov/27/...
Storing times for human events
I’ve worked on various event websites in the past, and one of the unintuitively difficult problems that inevitably comes up is the best way to store the time that an …
simonwillison.net
November 28, 2024 at 6:14 AM
Reposted by Sergi de Pablos
I think a more probably route would be a better, “DMZ”-like destination that is more useful than CSV.

Think, if most SaaS apps could write to Iceberg, and your CDW could read that iceberg .. each only writes 1 integration and everyone wins!
November 23, 2024 at 1:24 AM
Reposted by Sergi de Pablos
I am utterly delighted by this subversive use of the Bluesky label feature, where you can subscribe to a custom app that then adds visible labels to profiles

This one shows you which posh expensive private school British public figures attended, plus their current fees
bsky.app/profile/dadd...
bsky.app
November 22, 2024 at 3:53 PM
I've already seen some people offering their API documentation in a LLM prompt optimized format, like
docs.jina.ai
November 22, 2024 at 5:48 AM