Jonathan Halcrow

@bneyshabur Congrats and Good luck!

4) Building a company to build a technology to accelerate science Now I'm starting something new—focused on core bottlenecks that could unlock step-change acceleration across science and technology. It's ambitious. I'm about to learn a lot and be humbled. I see so many similarities between my experience in this new journey now and my weeks-long backpacking trips in the Alaskan wilderness with no guide and no trail. If this excites you and you want to learn more, reach out!

I've left Anthropic to start something new. 🧵

An exciting milestone for AI in science: Our C2S-Scale 27B foundation model, built with @Yale and based on Gemma, generated a novel hypothesis about cancer cellular behavior, which scientists experimentally validated in living cells.  With more preclinical and clinical tests, this discovery may reveal a promising new pathway for developing therapies to fight cancer.

@graph_ @suchenzang 👀

@suchenzang tbh it’s pretty obvious who is the strongest L8 in Gemini is

are L8s at deepmind actually strong? 🤔

Think your LLM is smart? Try Big-Bench EXTRA Hard. The real test. #Reasoning #Benchmark #LLM

While I'm at home with my newborn (👶🏻❤️), you can still catch my work at #NeurIPS🧵:

@graph_ Are you accusing me of stolen valor?

@jhalcrow you're supposed to post your own, followed by the worst career advice imaginable

Academics the microsecond they get the chance to post their h-index on twitter

Think your LLM is a time wizard? 🧙 One might think they answer temporal questions through reasoning, but under the hood they may solely rely on their parametric knowledge 🤯 Our new benchmark, ToT, is specifically designed to measure the temporal reasoning capabilities of LLMs.

🚨 Transformer theory alert 🚨 What algorithms can transformers execute efficiently? Our preprint sheds some light on reasoning capabilities of transformers, now in 𝒓𝒆𝒂𝒍𝒊𝒔𝒕𝒊𝒄 𝒑𝒂𝒓𝒂𝒎𝒆𝒕𝒆𝒓 𝒓𝒆𝒈𝒊𝒎𝒆𝒔. Paper: arxiv.org/abs/2405.18512 More in thread! 1/8

↪️ It's "Talk Like a Graph" day at #iclr2024 - 3:15, at #google see a *live demo* of GraphToken, where we train a tiny model (Gemma 2B) and get sota! arxiv.org/abs/2402.05862 - 4:30, stop by poster 89 to see the TLG leaderboard and talk with us! research.google/blog/talk-like…

We'll be demoing some of our work on graphs and LLMs later today at ICLR. If you are around stop by and say Hey!

@SaffronOlive Etrata!

Let's say you were building Felix Five-Boots, what's one card that you are definitely adding to the deck?

@agihippo Beyond the physical system models like GraphCast (from GDM) we have a lot of impact inside Google with anti-abuse applications. Graph representations are really effective in that domain in a way that's hard to convey with public datasets / open models

what happened to graph neural networks? do people still care about it? it feels fashionable to get an easy neurips paper or so from it but what impact as GNNs made? serious question.

@graph_ Presubmit failed - Unused import

Definitely not PEP8, 2/10

Talk Like a Graph or Let Your Graph Do the Talking 🧠🕸️ New paper alert 🚨: we show how GNNs can effectively encode graph-structured data for LLMs: arxiv.org/abs/2402.05862

𝐋𝐞𝐭 𝐘𝐨𝐮𝐫 𝐆𝐫𝐚𝐩𝐡 𝐃𝐨 𝐭𝐡𝐞 𝐓𝐚𝐥𝐤𝐢𝐧𝐠: 𝐄𝐧𝐜𝐨𝐝𝐢𝐧𝐠 𝐒𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞𝐝 𝐃𝐚𝐭𝐚 𝐟𝐨𝐫 𝐋𝐋𝐌𝐬 Don't know what to do with your graphs in 2024? Shove them to an LLM, of course, and let LLM figure out what to do! arxiv.org/abs/2402.05862 Short thread (1/5):

New paper alert 🚨: "Talk Like a Graph" studies the problem of encoding graph-structured data, paving the way for AI to better understand and process complex relationships. Surprising results below 🤯! #AI #LLM #GraphData #MachineLearning

@mtkonczal This is kind of a dark guess, but funeral expenses?

Trivia: BLS tracks 315 CPI categories across sectors/totals. Running own-item Phillips Curves on each for 21C, only 1 has stat-significant unemployment coefficient = -1; or its inflation moves 1-to-1 with the economy. What is it? Hint: has a small weight and it'll make you mad.

FYI: we're running a Kaggle contest with GNNs!

Packed room for TFGNN tutorial today at #KDD2023!