Marc Brockschmidt

@BlancheMinerva @JeffDean The license text says "If, on the Llama 2 version release date, the monthly active users [...] is greater than 700 million monthly active users [...] you are not authorized to exercise any of the rights under this Agreement" Can you explain how this matches your statements?

See the blog and tech report for more details: blog.google/technology/dev… goo.gle/GemmaReport The entire Gemini/Gemma team are really excited to see what people do with these models. We think they are a great tool for a wide variety of use cases!

@josepablocam @miltos1 Indeed, we had a standard wrapper for main methods doing exactly this in our dpu-utils collection. The only problem is that @miltos1 joined MSR after Jack was my intern, and I'm now very confused. Debugging time travel?

Ed Balls. Sorry, I mean, @mmjb@sigmoid.social.

@schemeprincess I've made good experiences with making sours using Lyre's non-alcoholic spirits (particular the campari analogues). The acidity from the lemon really helps to substitute for the kick from the alcohol.

Tweeps: what are good non-alcoholic drinks? I want to drink something other than flavored sparkling warmer but I don’t like drinks that are too sweet and prefer bitter drinks. This is apparently really hard to find in non-alcoholic form!

In programs such hyperedges appear naturally all the time - the token sequence, function calls, data flow, etc. In the paper we show how our model can exploit this to outperform strong GNN and Transformer baselines. 3/3

The intuition is that the concept of a qualified hyperedge (e.g. student(name=..., major=..., ...)) can be treated like a sequence in a Transformer, with qualifiers as "positions". That yields a "message" for each adjacent vertex, and you can treat that like message passing. 2/3

Some of my last work at MSR, led by @DobrikG and @miltos1. It introduces a transformer variant capable of processing hypergraphs, and shows how that is useful for program and knowledge base tasks. 1/3

@eaftandilian @satnam6502 Still love remembering that time going to NeurIPS when the two dudes in the row in front of me worked out that they both work for DeepMind after seeing each other's sets of laptop stickers after takeoff.

@satnam6502 That’s my favorite game when traveling to conferences: spot my fellow nerds

Pretty sure I have positively ID-ed every passenger on this EasyJet flight from Gatwick to Ljubljana who is a functional programmer. From hair styles, t-shirts and overhead conversation about Agda it is not very hard to spot the tells. #ICFP2022

@lorisdanto To be fair, 8-year olds are most likely to share your philosophy, right?

Friend: "we're working hard interviewing sitters to find one that has our same philosophy etc" Me, who's just paid an 8-year old $10 to hang out with my 3-year old:

Sweet news: today, I'm starting as a Research Scientist at @GoogleAI, working with old friends such as @dtarlow2, @RandomlyWalking and @miltos1 on learning to assist software engineers. The team, the data, and the opportunies for impact are truly exciting!

Bittersweet news: a week ago, I had my last day at Microsoft Research. I've spent almost 9 years there, and learned so much from the lovely and brilliant people working at MSR. A new adventure is coming, but first, I'll enjoy 3 months of a break.

@awsTO To be fair, most editors of British newspapers went with that photo as well...

@simonbatzner @jhrmnn @PetarV_93 Exactly (this is actually quite elegant with numpy-ish indexing and .add/etc). In the sparse variant (batch of graphs = 1 graph with many unconnected components) the number of nodes and edges obviously differs between batches. This is where in TF2 I use symbolic sizes.

@jhrmnn @mmjb86 @PetarV_93 Wouldn’t the shape still be dynamic?

#JAX & #GNN question: how do you make it go fast without padding? I ran into the jit.compile-caches-on-shapes issue, and it's not clear how to sidestep that. (Yes, I know that there's libs like jraph, but I'm just fooling around to learn)

@PetarV_93 Huh. So ... Why? I'm surprised by this, as XLA doesn't have the full static shape requirement (which I've heavily abused in the past with TF2's `tf.function(input_signature=...)`. Which step requires the static shapes?

@mmjb86 AFAICT, even libraries like Jraph rely on padding. The specific trick commonly used is to pad to the nearest power of two, which reduces the amount of distinct shapes that would need to be jitted while not making the memory usage too wasteful.

@gweissenbacher @flxflx I should have organised a proper betting pool, now that you bring it up

@lorisdanto @AI4Code As many as it takes to get to average human-level expertise. So ... Roughly 25, I guess

@mmjb86 @AI4Code But how many neurons?

@AI4Code More code, fewer molecules is the plan, so watch out :-)

@mmjb86 I greatly enjoyed reading the work you did at MSR. Good luck with your next gig!

@foo_fighterin Thanks! I should say that you and everyone else at Novartis were extremely kind to this bumbling neophyte. I'm very thankful for your patience and generosity in teaching me about drug discovery.

@mmjb86 Very sad to hear this, it was really great working with you, I learned at lot! Wish you all the best for the new adventure! Enjoy your break!

@flxflx Suspense 😉

@mmjb86 What's next?