Brett Tully

Geospatial AI | Numerical Science at Scale | MLOps.

Snr Director, AI Output Systems @nearmap.

1st employee @FLFusion.

[Brett Tully @brett_tully]

@opensauceAI @Wing @TheSlowMoGuys There aren't many things in the world that don't look better at 200k frames per second. This was a very cool episode!

[Brett Tully @brett_tully]

@TobyWalsh @nathanbenaich @airstreet @stateofai .@nathanbenaich you might want to fix your socials thumbnail, looks like the report is a year old!

[Brett Tully @brett_tully]

Surprised that ProCan from @CMRI_AUS didn't get a mention in this review @EricTopol. While not looking at plasma, tissue proteomics has massive potential in cancer clinical workflows.

https://t.co/wBijKSrgym
https://t.co/kItH7TTwTc

[Eric Topol @EricTopol]

My piece @ScienceMagazine today
What happens when you can assay over 10,000 proteins in the blood, in tens of thousands of people with years of follow-up, and use #AI analytics?
An exciting, but largely unrecognized new frontier in medicine; open-access
https://t.co/LGkHiPRCH1

[Brett Tully @brett_tully]

Interesting!

[Nick McGreivy @NMcGreivy]

Our new paper in @NatMachIntell tells a story about how, and why, ML methods for solving PDEs do not work as well as advertised.

We find that two reproducibility issues are widespread. As a result, we conclude that ML-for-PDE solving has reached overly optimistic conclusions.

[Brett Tully @brett_tully]

Great sessions at the Qld AI Summit today by @_sholtodouglas @opensauceAI and others #qldaisummit

[Brett Tully @brett_tully]

@JonKrohnLearns @WebSummit Amazing mate, congrats!

[Brett Tully @brett_tully]

@hugobowne Fellow Aussie @maosbot has done a heap in this space.

[Brett Tully @brett_tully]

Genius. "Don't publish slop"

[Simon Willison @simonw]

Slop is the new name for unwanted AI-generated content https://t.co/4PXWNEYl7O

[Brett Tully @brett_tully]

@maosbot https://t.co/Isz6dtL6hi

[Brett Tully @brett_tully]

@EconAndrew To group the data like this, the categories need to align, but my experience of the two countries is that they don't. I also wonder if assessing pedestrian safety as part of the safety of the car might be part of it.

Tho, there are a lot of Ford Rangers on the road now!

[Brett Tully @brett_tully]

@TobyWalsh I certainly wouldn’t be making such definitive claims with research like this floating about: https://t.co/8HQesvQ8iX which, as all things frontier AI, is a dual use technology as shown in https://t.co/cEhTRwxohB

H/t @emollick for the links

[InceptiveMind @Inceptive_Mind]

It is the first privately funded fusion company to fire a shot on the Sandia National Labs' Z Machine.
@FLFusion #NuclearEnergy #fusion

https://t.co/VMEGKx6atw

[Brett Tully @brett_tully]

This is my kind of 🤓 article

@jjbw did you have something to do with it?

@HDKaul you'll like this one

https://t.co/04dnSPysFE

[Brett Tully @brett_tully]

@charliermarsh How is this different from the rip rust package from the @prefix_dev team? https://t.co/A7HL7JUcaT

[Brett Tully @brett_tully]

Pretty good list to end the year with

[Sam Altman @sama]

what i wish someone had told me:

https://t.co/1nEaYimzXG

[Brett Tully @brett_tully]

@prefix_dev This looks very cool!

It also looks like it's on the path for something I'd love to see in pixi: take a collection of meta.yaml / env.yaml files, solve the common dependency tree, export that as a conda metapackage that can be used to join the dependency tree across libs.

[Brett Tully @brett_tully]

@jeremyphoward @geoffreyhinton @ylecun Software much harder than hardware https://t.co/kkWGFgsMhc

2/2

[Brett Tully @brett_tully]

@jeremyphoward @geoffreyhinton @ylecun It may not be a direct analogy, but it seems inevitable that at some point in the near future, models will be thought of as dual-use. If you believe that, then it follows, open-source will need to be purposefully hobbled (c.f. EOS simulations).

1/2

[Brett Tully @brett_tully]

I asked DALL-E 3 to make a social media post for the @brisbaneroar for the Aus Cup final. We've come a decent way from the first avocado chair in a pretty short time!

[Brett Tully @brett_tully]

And this for vision

https://t.co/7m4ZowHA9o

[NirSD @nirsd]

7/n what information do these extra registers store?
Holier shit, its totally coherent concepts that we can understand, the cup, the spoon, the tea all picked out as salient individual features by the model without being asked to.

[Brett Tully @brett_tully]

This is absolutely mind boggling. I sense we are on the cusp of a step change in technology. Today's use of models will seem quant in a few years time.

[Anthropic @AnthropicAI]

The fact that most individual neurons are uninterpretable presents a serious roadblock to a mechanistic understanding of language models. We demonstrate a method for decomposing groups of neurons into interpretable features with the potential to move past that roadblock.