Ive hardly kept this quiet but very excited/other emotions to say we’re building
@finsterai
! Finance meets GenAI meets other things. Lots to do, lots of fun, we’re hiring! Stay tuned. As they say, show don’t tell etc. Danke!
@CharlotteCGill
I mean this is actually useful history and thats funding over 5 years. That's the equivalent of paying 1.5 junior analysts at a bank to do actual historical research of a 500 year period and build a wealth of history of our country. Not everything you dislike is woke.
@CharlotteCGill
The “anti woke” argument is that we should preserve our culture and heritage and not attack institutions/history is something we should be proud of. How do you do that if you dont study your history, culture and heritage?
@CharlotteCGill
The free market wont fund this. So it has to be funding bodies. This is all basic social and economic policy that regardless of whether you lean left or right is obvious. But in the name of culture war, we will dismantle academic institutions just to “own the libs”.
DL twitter is really something these days because you just see people constantly fighting about whether scaling is enough while the world has far bigger problems but as a cherry on top neither of the people do any real research anymore and last wrote code in 2012
Every few months I remember my idea for a “machine learning at the pub” podcast where as the evening progresses the ML takes get spicier. It’d be called WhiskAI and my first guest would be
@egrefen
😂
Hola — if you’re based in LON/NYC (or well anywhere) — I’m hiring founding engineers/AI folks for a ~new thing~. The key bits are: we’re moving quickly, Im super excited by it, minimal BS, zero techbro-ness and well, bloody good fun. And shit banter. DM me! Grazie
Really excited about our latest work showing that large Transformer-XLs can be used in RL agents. We show SoTA performance on DMLab with gated transformers and a few small changes. Led by Emilio as an internship project!
@DeepMindAI
Finally, Transformers working for RL! Two simple modifications: move layer-norm and add gating creates GTrXL: an incredibly stable and effective architecture for integrating experience through time in RL.
Great work from Emilio interning at
@DeepMindAI
A beautiful parting gift to receive from friends at
@GoogleDeepMind
! Im super touched! Combining my go to passions of cricket and AI😂
not enough Virat Kohli puns on here but huge thank you to everyone! It’s been an honour!
A quick update from me! After 7 wonderful years of meeting great folks, my terrible banter, generally mucking about and some research to boot, today was my last day
@GoogleDeepMind
. Super grateful for the time of my life, but time for a wee break for me. Ill go back to tweeting
Really happy and a lovely early-Christmas present that our paper "Multiplicative Interactions and Where to Find Them" was accepted to
#IClR2020
. We analyse such interactions, unify different forms (eg hypernets, gating), and encourage you to use more of them! 1/2
The only reason we have this level of AI right now is cause people worked on it even when it wasn't cool and funding was dry.
Work on whatever makes you happy, whatever gives purpose, whatever you cant not work on, whatever pays the bills, whatever you want.
Happy to share that our paper "Top-KAST: Top-K Always Sparse Training", proposing a new sparse training method, was accepted to NeurIPS 2020!
This was work done at
@DeepMind
with
@rpascanu
@sindero
Jack Rae and
@erich_elsen
!
1/3
Happy to share this blogpost about our work on compressive transformers (led by Jack Rae)! We show that you can extend the effective memory of Transformers(-XL) with a simple compression scheme and show gains for LM and RL! (, accepted to ICLR 2020)
Memory is a crucial feature of intelligence. Our new blog post overviews the use of memory in deep learning, and how modelling language may be an ideal task for developing better memory architectures:
I have a free
@iclr_conf
registration as a reviewer. I’d like to effectively pass this on by paying for someone’s registration who otherwise may not be able to attend. Please DM if interested or feel free to retweet :)
Our new paper out on Arxiv: "Distilling Policy Distillation"! We investigate policy distillation for RL: how to use it, what form to use and show some nice theoretical results on the way. Accepted as an Oral at AISTATS :)
@wojczarnecki
@maxjaderberg
@rpascanu
@sindero
@DeepMindAI
Unrelated to the board decision, but I love this “started in apartment” meme because it lets you cosplay as “coming from nothing” even when thats not the case. (He was CTO of Stripe before OpenAI)
Congratulations
@demishassabis
what an honour! Demis saw AI coming way back when. And also personally why I got into AI. Met Demis
@QueensCam
CompSci dinner in 2014 and was floored at the Atari demo; hounded DM to let me intern, dropped out of Uni and stayed on. Very grateful
New paper with
@janexwang
and others
@DeepMindAI
on Meta-Learning in the presence of recurring tasks! Accepted to
#icml2018
:)
"Been There, Done That: Meta-Learning with Episodic Recall"
1) CONGRATULATIONS to everyone at DeepMind; this is AMAZING!!! ❤️
2) Let this be a lesson to every techbro who blows with the wind and tweeted "Google is dead" for the last 6 months. Ignore the noise.
3) Stop confusing innovators dilemma with "Google can't ship". They're the
The Gemini era is here. Thrilled to launch Gemini 1.0, our most capable & general AI model. Built to be natively multimodal, it can understand many types of info. Efficient & flexible, it comes in 3 sizes each best-in-class & optimized for different uses
@aidangomezzz
Imo the real worry is if the loss function suddenly spikes UP. Maybe the model is using morse to speak to us? Maybe it’s getting worse on purpose??? Makes you think. Very worrying. About to cancel some runs
On the
@NipsConference
poll: it's hard to get to the point where we're an inclusive community if we're going to follow the results of a male-dominated sample voting on an issue that female/non-binary respondents (and a good chunk of male ones too) found offensive.
#ProtestNIPS
This is how all startup projections sound. “If we even get half of Googles users, and also we make them pay for email, then we will be at $25B revenue”. Thanks mate
Ran some numbers on this. Gmail apparently has 1.8 billion users. If Elon launches XMail as part of the Basic $3/mo subscription and gets just half of Gmail users to switch over that’s $2.7 BILLION A MONTH in new revenues for X. Wow!
@MohapatraHemant
@perplexity_ai
Imo irrelevant that theyre 2 years old. Theyre well funded, have some of the most well known advisors and investors and worth close to what the NYT is market cap wise. Cant hold tech startups to diff standards just because things move fast imo; regardless of ones views on this
It's been well...quite a year since my Danish Christmas last year. Solo virtual Christmas in London wasn't quite the same but I tried to Scandi it up -- here's wishing you all a very merry Christmas and I hope 2021 is all you want it to be!
@sriramk
@k0ol1
Think that's a bit uncharitable Sriram -- there's loads of people doing a lot on A. Both are possible and there's a lot of people working on it. We're the ones who left, we hardly get to critique people who bring up A; it's not really whataboutism to point out structural issues
Dishoom's net contribution to the development of AI cannot be understated. It was our personal pilgrimage site at DM and we've kept it up
@finsterai
for good measure. I think the chilli cheese toast brings out creativity. It's all the ChaatBots.
@freyaindiaa
It's not missing though is it -- most mental health advice will tell you about volunteering and doing things for others and alot of therapy is trying to get you to be self aware about why you did something and why you felt something and looking at other people's perspectives.
Virtually nobody is pricing in what's coming in AI.
I wrote an essay series on the AGI strategic picture: from the trendlines in deep learning and counting the OOMs, to the international situation and The Project.
SITUATIONAL AWARENESS: The Decade Ahead
If you want to hear more about our NeurIPS submission, we'll be at poster session 7 tomorrow (a very early 5am GMT for me but it times well with the Australian cricket summer)!
I wish this was a subtweet to someone in particular but sadly it’s not: scrolling through some AI startup team pages and it’s like people forgot you can hire women. Just dudes in a room with a careers page talking about how open and inclusive the culture is. You can basically
One of the coolest results (benchmark-wise) I've seen recently is TransformerXL (30-ish ppl to 18 on Wiki103!) this ICLR. BigGAN's been talked about loads and despite the surprisingly average reviews TransformerXLs got -- yet another win for attention based models! Really cool
Excited to be in New Orleans for
#ICLR2019
! I'll be at the
@DeepMindAI
stand tomorrow to chat about research in general and the research engineering roles!
Not doing the whole PR thing as yet, but wanted to give a shout out to amazing investors/angels who backed
@finsterai
and me, a first time/solo founder. We're rolling our to our early customers and continue building/hiring. 1/2
My yearly Christmas thread. For all those in quarantine, take care and hold on! Have personally been lucky to have my first Christmas back in India for 2 years. My Mum has also happened to take this very impromptu but Christmas-card looking photo of me. Take care you all!
It's been well...quite a year since my Danish Christmas last year. Solo virtual Christmas in London wasn't quite the same but I tried to Scandi it up -- here's wishing you all a very merry Christmas and I hope 2021 is all you want it to be!
You can make up anything you want on Twitter. Spent 7 years at DeepMind, I can assure you that tuning models is very hard and despite twitter imagination, there is no “wokeness” hyper-parameter. Not everything is a culture war. We have better things to do.
Looking at my social media, the normies have now found the racist Gemini images.
Their assumption is this must be a mistake by Google. But it's not. This was deliberate.
Like any religious or ideological cult, Google believes its mission is to convert the unbelievers.
Awesome to see this public! I didnt contribute at all to this while at DeepMind but def slowed the team down by requesting total garbage prompts on the early versions and that they were nice enough to try. My music taste generally flits between jazz and bollywood so you know.
Thrilled to share
#Lyria
, the world's most sophisticated AI music generation system. From just a text prompt Lyria produces compelling music & vocals. Also: building new Music AI tools for artists to amplify creativity in partnership w/YT & music industry
Given that every tweet on my timeline is now just:
“This thing has happened, here’s why you should care.
A 🧵 thread!”
I would like to state for the record. This tweet is not a thread and you don’t need to care. Merci.
Woke up to see this in
@sytaylor
’s newsletter re
@finsterai
! Thanks Simon!
Still early doors for us but am super excited about what Finster’s Agents can do for finance/capital markets. Let people focus on the important bits & let AI do goal directed research/analysis/comps etc
And thank you especially to those involved in this! Twitter seems especially bad at letting me tag anyone today but shout out to Jackie, Dushyant,
@egrefen
, Dan,
@tfgg2
, Matt, Yotam, Josh,
@pfau
@janexwang
and Raz!
Just took my first day off since founding
@finsterai
. Attempting to leave laptop away. On the beach. Day 1. Walking back to room. Two women talking about tech enabled services and disrupting insurance. GG WP.
I tweet random AI things all day for 3 likes and then the thing people actually read is when I ratio todays twitter main character (the goal of Twitter is to never be the main character of the day etc). Many ways etc
Two different people including a tech reporter mentioned today to me that they enjoy my social media presence. Unclear if this is very good or very bad. Will continue posting till it becomes clear. Onwards
Happy to share our latest ICML paper -- work with Wojtek,
@maxjaderberg
,
@sindero
,
@yeewhye
,
@lqh20
and others at
@DeepMindAI
! We train large actions spaces and models by introducing simple curricula for RL agents -- mixing of policies + distillation from simpler policies.
Our
@ICMLconf
paper introduces Mix&Match - a general-purpose framework for training complex RL agents. M&M creates curricula over the agent, using solutions found by simpler ones to train harder-to-train ones
NEW WORKSHOP: Sparsity in Neural Networks: Advancing Understanding and Practice (July 8-9, 2021). This workshop will bring together members of the many communities working on neural network sparsity to share their perspectives and the latest cutting-edge research (Deadline: 6/15)
Real innovation happens slowly, and then all at once. Easy to move very fast after decades of slow, meticulous work. We stand on the shoulders of giants etc.
Quick self indulgent tweet: Given we’ve done p little branding/pr (so far), it’s always nice to see
@finsterai
turn up on a list this early :)
Tis customers, product, revenue in that order of prioritisation so far.
stay tuned for lots of updates in the summer hopefully!
AI agents are one of the hottest areas of development & investment in AI
Here's a deep dive into the ecosystem from
@Prosus_Ventures
:
1. AI agents + AgentOps landscape
We really should teach more people stats and Bayes. My man here forgot the p(dem|city) term. This is like saying most failures of Ford cars are due to Ford and not Hyundai. Like yes.
5 years later, and NeurIPS rolls off the tongue just fine and Id totally forgotten that there even was a name change. See, it’s not hard to change things sometimes. Even in academia. Is good.
I also started
@finsterai
in my apartment. And was unemployed working on it for 4 whole months! **spent 7 years at DeepMind before, had savings, generally privileged, went to Cambridge. SO MUCH HUSTLE
Thank you to
@Siftedeu
for featuring one smart and one dumb quote from me lol
Good read from Tim Smith here
Am here so I can go to Lords more btw but genetically sworn to Liverpool. (Also congratulations to Arsenal and Rory; good week for the UK)
@MohapatraHemant
Great question! An interesting variation is to say "fiction" *and* released in the 21st century to exclude James Joyce or any non fiction techinical-adjacent books like GEB haha
Merry Christmas to everyone from Bombay! Escaped rainy London to celebrate the holidays, attend my closest friends weddings, and also spend time with my new best mate. Stay safe, and may 2023 bring you lots of happiness, even larger LMs and most of all, more Twitter drama xx
My yearly Christmas thread. For all those in quarantine, take care and hold on! Have personally been lucky to have my first Christmas back in India for 2 years. My Mum has also happened to take this very impromptu but Christmas-card looking photo of me. Take care you all!
Need a word for that sweet spot that AI products need to be at in-between copilots and backseat drivers.
Like your friend who gets control of the tunes on a road trip and only plays bangers
Of course things are getting better but to sum up the state of the theory in Deep Learning: there’s a famous
@karpathy
tweet about 5e-4 being the best learning rate and while 100percent irony it was not entirely terrible advice as a first choice😂
Apparently there are people out there that think deep learning is too theory obsessed
I can't even rub two propositions together to create a third
Guess I'll try and become a more prominent researcher, for their sake
I think a lot of AI Hot Takes can be better understood if you take into account that 1) Something about Twitter makes everyone go MAD
2)People who take the time to post their Hot Takes live in the biggest bubble/ echo chamber of all time
3) Careful nuanced debate is boring
Can't +1 this point enough. Just as most criticism of deep learning is instead criticism of "pure DL will not get us to AGI" -- but I don't know a single person doing DL who thinks that in that first place.
@ylecun
@hardmaru
Bit of a pedantic point, but I think we should instead criticize model-free RL rather than "pure RL" (not sure what this is). We're still doing RL-based learning with an internal model.
Underrated benefit of being a CEO is I can use
@finsterai
's account to like my own posts; providing instant dopamine when I get a notification on my phone 4 minutes later and forget that I liked my own post.
My two favourite papers I've written are Multiplicative Interactions and Where to Find Them and TopKAST (also like their titles fwiw); one was inspired by and built on HyperNetworks and the other by evolutionary approaches to sparse networks. Basically I love
@hardmaru
's work
Introducing Evolutionary Model Merge: A new approach bringing us closer to automating foundation model development. We use evolution to find great ways of combining open-source models, building new powerful foundation models with user-specified abilities!
Gemini 1.5 Pro - A highly capable multimodal model with a 10M token context length
Today we are releasing the first demonstrations of the capabilities of the Gemini 1.5 series, with the Gemini 1.5 Pro model. One of the key differentiators of this model is its incredibly long
Glad we have managed to replicate the awkwardness of real poster sessions virtually 😂
#NeurIPS2020
Particular fan of the realism of the virtual drive by
I know this is old news but I continue to be baffled by the fan following of Twitter “thought leaders” whose basic claim to credibility is they’re smart and made a lot of money being in Tech. And this somehow allows opining on complex social policy
It’s interesting that the community spent a long time adding in memory/planning etc to Deep RL agents only for it look like end-to-end always won and now we have a new generation of people adding memory, planning etc to LLM-based agents…jury still out?
Cant quite believe what Ive just seen! My first time away from home for the winter, but it’s been seriously enriched by having witnessed one of the greatest test series. Tearing up at this unbelievable Indian comeback. What an inspiration - cap is very much doffed!
#GabbaTest
We've truly come a long way from "don't worry about AI safety, people will be careful" to "we must fight the woke AI"...
Living in a very dumb timeline
A huge plus that no one told me about leaving full time employment and doing your own thing is that you can work at 7am or 1am but you can also crucially, take a nap at 1pm, and like that’s amazing. Also you can do all this while dog sitting. The lil one also naps.
Respectfully disagree - very much not a red flag. The line between ML Eng and Research is blurry and in theory you dont need any degree for either (though helpful). AI’s had a gatekeeping problem for ages and also a reverse gatekeeping problem (experience bad, indie hacker good)
Today, we're announcing Claude 3, our next generation of AI models.
The three state-of-the-art models—Claude 3 Opus, Claude 3 Sonnet, and Claude 3 Haiku—set new industry benchmarks across reasoning, math, coding, multilingual understanding, and vision.