Lightnews — Scholar-powered news

Joe Hellerstein @joehellerstein.bsky.social · Jun 26

Reposted by Joe Hellerstein

Madelon Hulsebos @madelonhulsebos.bsky.social · Jun 5

Let’s collaborate on democratizing insights from tabular data in Amsterdam! ✨

PhD directions: 1) fundamental techniques for tabular foundation models, 2) reliable mechanisms for AI-powered tabular data analysis.

Sharing w/ friends appreciated! ⬇️

ELLIS @ellis.eu · Jun 5

🏹 Job alert: 2 fully-funded PhD Positions at Table Representation Learning Lab - @ellisamsterdam.bsky.social‬

📍 Amsterdam 🇳🇱
📅 Apply by June 30
🔗 More info: https://bit.ly/4519pj1

Open positions | TRL Lab

bit.ly

1 3 2

Joe Hellerstein @joehellerstein.bsky.social · Jun 5

The last blog post in my miniseries on CRDTs is up!

jhellerstein.github.io/blog/crdt-in...

Mix of pragmatism and formalism.

There's actually a small result in there that may be novel: Strong Eventual Consistency !=> Determinism. Curious to hear whether they've seen this result elsewhere.

CRDTs #4: Convergence, Determinism, Lower Bounds and Inflation

The CRDT literature sometimes leaves room for mathematical ambiguity. Maybe because the bulk of the work tends to be targeted at systems researchers and…

jhellerstein.github.io

2 13

Joe Hellerstein @joehellerstein.bsky.social · May 28

posted today!

BTW I peeked at the automerge Rust? Collaborative editing is an example where one probably *has* to resort to unsafe behavior (you're the expert there!) so I'm mostly advocating for more encapsulation/comments in that case.

jhellerstein.github.io/blog/crdt-do...

CRDTs #3: Do Not Read!

Ever used a CRDT, thought you were safe, and—boom—you bought a Ferrari you didn't mean to? It could happen to you! The truth is that CRDTs are dangerous to…

jhellerstein.github.io

4 6

Joe Hellerstein @joehellerstein.bsky.social · May 28

Next blog post in the CRDT Series is up!

This one is for the developers... stay safe out there, folks.

jhellerstein.github.io/blog/crdt-do...

CRDTs #3: Do Not Read!

Ever used a CRDT, thought you were safe, and—boom—you bought a Ferrari you didn't mean to? It could happen to you! The truth is that CRDTs are dangerous to…

jhellerstein.github.io

1 2 14

Joe Hellerstein @joehellerstein.bsky.social · May 28

Good thread. Thoughtful as always.

Charity Majors @charity.wtf · May 26

I have not paid a ton of attention to the uproar over RTO policies, bc we are all in on distributed teams and not going back.

My impression (via social media) has been that these were shadow layoffs.

Last month I asked an investor why they are doing RTO. He said: "Retention, mostly. And morale."

3

Joe Hellerstein @joehellerstein.bsky.social · May 25

Really early and well seen, definitely influenced me and my team! Hats off.

1

Joe Hellerstein @joehellerstein.bsky.social · May 25

Depends what you want the “set of lists” semantics to mean. I’d think you likely want a 2P-map lattice of RGAs (2P-map would be like a 2P-set but with a lattice value associated with each unique item in adds). If you want more detail please comment in the blog so it’s easier for others to find it.

Joe Hellerstein @joehellerstein.bsky.social · May 25

There are simple and helpful composites that can be written generically and reused safely. E.g. lattice pairs (free or lexical) and Map lattices. Helps to have a language with good support for generics (parameterized types).

1

Joe Hellerstein @joehellerstein.bsky.social · May 22

(Catching up to my LI feed).

Next blog post is out! This is the first real post in a short series on CRDTs, an idea that has some currency in the distributed programming community, but one that comes with a number of sharp edges. Be careful out there!

jhellerstein.github.io/blog/crdt-tu...

CRDTs #1: Turtles All the Way Down

This is the 1st post in a series of 4 detailed posts I'm doing on CRDTs. Please see the intro post for context. Modern distributed systems often seem to rest on…

jhellerstein.github.io

1 3 12

Joe Hellerstein @joehellerstein.bsky.social · May 22

Blog relaunch! Bbye wordpress, hello github.

If you're into SW dev, cloud, databases, distributed systems, automatic codegen ... or data and CS in general... check it out.

As a warmup, I'm starting with a series of posts on CRDTs. Intro post up now: jhellerstein.github.io/blog/crdt-in...

A Run of CRDT Posts

Over the next few days, I'm going to post a number of observations about CRDTs: Convergent Replicated Data Types. These are data structures that aspire to help…

jhellerstein.github.io

2 5 18

Joe Hellerstein @joehellerstein.bsky.social · Apr 16

Wow! @arvind.bsky.social giving an awesome keynote including discussion of VegaExpress and GoFish interactive vis libraries from his group. #EPICRetreat #UCBerkeley.

2

Joe Hellerstein @joehellerstein.bsky.social · Apr 16

Here’s a provocative example from JD Zamfirescu-Pereira on ways that humans and LLMs can get misaligned on expectations. Is the LLM lying? Is it just emitting tokens? How do people interpret this? #EPICRetreat #UCBerkeley.

1 2

Reposted by Joe Hellerstein

Shadaj Laddad @shadaj.me · Feb 13

The SF Systems Meetup is back! On 2/27, we're excited to have headline talks from the creator of FizzBee and a research collaborator with Signal. This is going to be a super fun night diving deep into making distributed protocols work, hope you'll join us! lu.ma/vqjf30k3

SF Systems Meetup: Correctness and Security for Distributed Systems · Luma

The SF Systems Meetup is back for the new year! This meetup, our theme is correctness and security. It's easy to write a distributed protocol, but very hard to…

lu.ma

3 5

Joe Hellerstein @joehellerstein.bsky.social · Feb 13

GPT4o did better:

GPT4o shows that f(a,b) = (a+b)/2 is an example of a commutative function that is not associative.

1

Joe Hellerstein @joehellerstein.bsky.social · Feb 13

The question: "what are examples of commutative functions that are not associative?"

GPT4 was funny, thinking aloud and then proving itself wrong:

GPT4 asserts that Min and Max functions are commutative but not associative, but then checks itself and backtracks.

1

Joe Hellerstein @joehellerstein.bsky.social · Feb 13

In some kind of sad watershed, today was the day as a professor when I live-ChatGPT'ed the answer to a question in a Zoom with my PhD student and his undergrad mentees.

But hey, let's paint it in a positive light: this was a demonstration of using the right tool at the right time.

1 5

Reposted by Joe Hellerstein

Eric Colson @ecolson.bsky.social · Feb 6

Operationalizing Machine Learning: An Interview Study by @joehellerstein.bsky.social, @adityagp.bsky.social, et al. Particularly love the part on "Retrofitting Explanations".
#MachineLearning #MLOps #Datascience.
arxiv.org/pdf/2209.09125

1 4 13

Joe Hellerstein @joehellerstein.bsky.social · Jan 28

I think “getting all of your coordination under one roof” (or behind a unified api or something) is the message I’m hearing from you. Don’t know if that helps?

Joe Hellerstein @joehellerstein.bsky.social · Jan 24

A muddled post at best. A sequential log *is* a point of coordination. It doesn't avoid coordination as claimed, it just centralizes it in 1 service (and arguably encourages overuse). Coordination avoidance is orthogonal: discover when global ordering is not needed. Ie avoidance avoids the log!

2 8

Joe Hellerstein @joehellerstein.bsky.social · Jan 23

Sunset in #Berkeley these days is a perfect field goal over the golden gate bridge. Shifts quite a ways north during the summer.

9

Joe Hellerstein @joehellerstein.bsky.social · Jan 8

2025. What a time to be alive!

"Whats new in Excel" dialog box. The text says "Data Aggregation Functions: We've added two incredibly powerful new data aggregation functions: GROUPBY and PIVOTBY"

1 9

Joe Hellerstein @joehellerstein.bsky.social · Jan 8

Fickle faculty followup follies

2

Reposted by Joe Hellerstein

Jamie McCarty @jamiemccarty.bsky.social · Dec 30

It’s incredibly beautiful that President Carter is our emissary on a Voyager probe. His words live on across our galaxy!

8 42 190