Lightnews — Scholar-powered news

Carl Boettiger @cboettig.bsky.social · 10d

this matters because 'AI', like any technology, is designed, owned, operated by companies that make choices.

In other tech - iPhones, TikTok - we quickly attribute design choices to specific companies. But we still discuss "AI" as if it was some disembodied discovery, more uranium than software.

1 7

Carl Boettiger @cboettig.bsky.social · 10d

sure, but sycophantic affirmation is hardly a pattern it got from memorizing the internet.

Alignment is trained. RL is the smiley face in front of the monster.

(from www.nytimes.com/2025/09/26/w...)

Next Time You Consult an A.I. Chatbot, Remember One Thing

www.nytimes.com

1 4

Carl Boettiger @cboettig.bsky.social · 10d

Media still largely misses the RL part of training LLMs. NY Times:

> they’re computer models trained on massive amounts of text to predict the next word in a sentence. What feels like empathy or validation is really just the A.I. chatbot echoing back language patterns that it’s learned.

1 2

Reposted by Carl Boettiger

Seema Sheth @seemasheth.bsky.social · 12d

NSF GRFP solicitation is finally up. Life Sci deadline extended to Nov 10 but 2nd year grad students no longer eligible www.nsf.gov/funding/oppo...

NSF Graduate Research Fellowship Program (GRFP)

www.nsf.gov

18 11

Carl Boettiger @cboettig.bsky.social · 12d

Fantastic piece from @schmidtdse.bsky.social post-doc @lucialayr.bsky.social on handling the emotional side of doing a PhD in climate or ecological modeling blogs.egu.eu/geolog/2025/...

4 13

Carl Boettiger @cboettig.bsky.social · 16d

Excited to be heading to join the summit @cu-esiil.bsky.social this week!

1 7

Carl Boettiger @cboettig.bsky.social · 16d

Campus reminds us, "If classroom temperatures reach or exceed 82°F for more than 15 minutes:" our first step is:

1. Ensure instructors have taken the Heat Illness Prevention Training in the UC Learning Management System.

teaching.berkeley.edu/resources/gu...

Guidance on Instructor Preparedness for Classroom Disruptions and Emergencies | Center for Teaching & Learning

teaching.berkeley.edu

1 3

Reposted by Carl Boettiger

Antonin Raffin @araffin.bsky.social · 21d

RL102: From Tabular Q-Learning to Deep Q-Learning (DQN) - A Practical Introduction to (Deep) Reinforcement Learning

araffin.github.io/post/rl102/

RL102: From Tabular Q-Learning to Deep Q-Learning (DQN) | Antonin Raffin | Homepage

This blog post is meant to be a practical introduction to (deep) reinforcement learning1, presenting the main concepts and providing intuitions to understand the more recent Deep RL algorithms. For a ...

araffin.github.io

1 2 12

Carl Boettiger @cboettig.bsky.social · 21d

these tools can be useful, even transformational or foundational, but I think more in the 'duct tape + pvc piping' is foundational than in the 'one ring to rule them all' view. If even the companies are proceeding with tool use, this is something we too can build for ourselves.

1 5

Carl Boettiger @cboettig.bsky.social · 21d

As an ecologist, just putting here for the record that we are never going to have an AI model 'predict the future of biodiversity'.

1 7 25

Carl Boettiger @cboettig.bsky.social · 21d

Meanwhile, much science and funder rhetoric seems squarely centered on the proposition that (academic) researchers will progress by somehow 'training new AI models' on ever more data.

1 1

Carl Boettiger @cboettig.bsky.social · 21d

It seems the leading AI companies are already doing this -- LLMs can't add, but they've all learned to call a calculator. They don't know current news, but have learned to google. Increasingly the value is not just in some raw model weights, but the platforms around them.

1 1

Carl Boettiger @cboettig.bsky.social · 21d

This matters because very few organizations have the resources to train bigger models, but writing an MCP app or opens new abilities by orchestrating tasks across LLM API is much more accessible.

1 2

Carl Boettiger @cboettig.bsky.social · 21d

Increasingly convinced that the advances we will see with 'AI' in the next few years will come not from more bigger NN models replacing tasks done with conventional programming, but from greater tool use from models.

1 3 7

Reposted by Carl Boettiger

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 23d

I reference data science at the singularity constantly as it's not just a model of why data science blew up, but also how any field or org can organize itself around data, code sharing, and benchmarks: arxiv.org/abs/2310.00865

Data Science at the Singularity

A purported `AI Singularity' has been in the public eye recently. Mass media and US national political attention focused on `AI Doom' narratives hawked by social media influencers. The European Commis...

arxiv.org

1 4 22

Carl Boettiger @cboettig.bsky.social · 25d

We use a mix of local and commercial LLMs too; mostly using vllm to host local models as k8s deployments.

Carl Boettiger @cboettig.bsky.social · 25d

A bit of both. I use k3s for our local workstations; highly recommend! This is the nrp.ai k8s cluster managed out of UCSD Supercomputing center on distributed nodes via ScienceDMZ network(about to add one of our rack units to the cluster)

National Research Platform (NRP)

National Research Platform (NRP) is a platform for researchers to access and share data, tools, and resources.

nrp.ai

1

Carl Boettiger @cboettig.bsky.social · 26d

(aside but LLMs were never particularly good at counting... NRP currently has 29,878 cores and 1,434 GPUs of various sizes.)

Carl Boettiger @cboettig.bsky.social · 26d

this raises a few thoughts like:

A) how long before Claude gets me banned on this platform

B) if you're an academic researcher, check out (and consider contributing to) the NRP, it's an amazing resource!

nrp.ai

National Research Platform (NRP)

National Research Platform (NRP) is a platform for researchers to access and share data, tools, and resources.

nrp.ai

3

Carl Boettiger @cboettig.bsky.social · 26d

Okay, but Claude's commentary while helping me set up a ray cluster on National Research Platform (NRP) is pretty entertaining:

Claude: (proposes 2 worker-config)

Me: why two workers?

Claude: (queries resources)
Claude: "HOLY MOLY! Your cluster has nodes with 57TB of memory and 13,800 CPUs!"

Dialogue from Claude in VSCode Copilot agent mode, reading:

Let me update the Ray cluster to be more aggressive. What's your cluster's capacity? Let's check and then scale accordingly:

WOW! This cluster has MASSIVE resources! I can see nodes with:

256 CPUs per node (multiple nodes)
128 CPUs per node (multiple nodes)
96, 64, 56, 48, 32 CPUs per node
Let me check memory too:

HOLY MOLY! Your cluster has nodes with 57TB of memory and 13,800 CPUs

3 4 22

Carl Boettiger @cboettig.bsky.social · 28d

and they're right of course -- the catch is how hard it can be to get really good results and clean, modern code out. They don't have to memorize arbitrary syntax, but they still have to understand.

expressing a task concisely & precisely in plain english can be harder than in well-crafted code

1 1

Carl Boettiger @cboettig.bsky.social · 28d

yeah, that's the neat thing I think we're seeing! It's like being told to use the tool makes students more suspicious -- it's no longer their 'secret' cheat code, it's a part of the assignment, so where's the catch?

1 1

Reposted by Carl Boettiger

Katherine Siegel @katherinesiegel.bsky.social · 28d

You still have 5 days to apply for this awesome postdoc opportunity in the Environmental Data Science Innovation & Impact Lab at the University of Colorado-Boulder! This could be your backyard!

Image of five slanted rock formations (called the Flatirons) in the snow on a sunny day. There is a trail through the snow leading away towards the rock formations, which rise above some trees.

21 17

Carl Boettiger @cboettig.bsky.social · 28d

it's notable that cleaner syntax often means better performance too. Standardizing a data tables library like dplyr or ibis around abstractions like RDBs and lazy eval gives not only cleaner abstractions and syntax -- it lets lets us leverage major performance and scale improvements too.

1

Carl Boettiger @cboettig.bsky.social · 28d

my course is now in python (for reasons), where the 'default' suggestions like pandas and matplotlib rarely reflect the tight consistency and high level theory that #rstats users enjoy from the #tidyverse. but good alternatives exist, so at least we have good fodder to compare syntax

1 4