Sign in with Twitter

Username:

Miles Brundage @Miles_Brundage San Francisco, CA

Research Scientist (Policy) at OpenAI. Research Associate, Future of Humanity Institute, University of Oxford. Views my own.

3,542 Following   24,123 Followers   43,203 Tweets

Joined Twitter 2/24/13


just noticed a bunch of #DEFCON27 talks are finally up! starting with watching @adversariel kill it https://t.co/kSasHyLkeE
Retweeted by Miles Brundage
11/17
2019
Bit biased but you should totally subscribe to my newsletter on the AI ecosystem in Europe. It’s written with love… https://t.co/lJrJArqa6K
Retweeted by Miles BrundageCheck out our latest effort to combat AI hype! Even well intentioned AI media coverage can get things wrong in a b… https://t.co/sxZf8Zxsh2
Retweeted by Miles BrundageBLOG | By fine-tuning an open source Transformer known as GPT-2, we can detect social media posts being leveraged i… https://t.co/mBZvPnyyyf
Retweeted by Miles Brundage
11/15
2019
@j_foerst Yup :) 🧠🤖🧠🤖"Learning From [Mouse] Brains How to Regularize Machines," Li et al.: https://t.co/3mPTHdLzR7 Wild stuff - they sh… https://t.co/38GrOz7grW
11/14
2019
Nice new results from @GoogleAI researchers on improving the state-of-the-art on ImageNet! "We...train a...model… https://t.co/NxSrjvYtpk
Retweeted by Miles Brundage
11/13
2019
Nice explanations of some key ideas in self-supervised learning by @lilianweng from OpenAI. https://t.co/7L8yyaDWlI
Retweeted by Miles Brundage
11/12
2019
A month ago, I was trying to demonstrate how to fix gender bias in contextual models like BERT for my book. I foun… https://t.co/juI8GVzBPp
Retweeted by Miles Brundage"Officers on the listserv are encouraged to adopt a Fight Club-style directive that precludes group members from di… https://t.co/ufrtadDjwj
Retweeted by Miles BrundageAs we shape our approach to synthetic and manipulated media, we think it's critical to consider global perspectives… https://t.co/DIwOtIEV0D
Retweeted by Miles BrundageUber didn't calibrate for all pedestrians, now a human is dead. **it wasn't a "mistake" - it was criminal negligen… https://t.co/iXPxtnumaS
Retweeted by Miles Brundage😊Self-supervised learning opens up a huge opportunity for better utilizing unlabelled data while learning in a supe… https://t.co/iujsAd41NM
Retweeted by Miles Brundage @gwern @togelius Not sure I parsed the console upgrade part? I was just looking at 2600. Yeah 1000 is short, though… https://t.co/SdNBDiQvcy @togelius I'd guess they're both in the trillions of frames ballpark, with NNs catching up or already passed. Rough… https://t.co/zRQWRwaDyb @evelyndouek The analogy between FB constraining itself with the OB and authoritarian governments introducing court… https://t.co/dwQ9UuYbIe"Facebook's 'Oversight Board:' Move Fast with Stable Infrastructure and Humility," @evelyndouek:… https://t.co/rJX7zO9SrK
11/11
2019
Canada is denying travel visas to AI researchers headed to NeurIPS — again https://t.co/O1zMBMxhzJ via @VentureBeat
Retweeted by Miles Brundage
11/10
2019
As nightfall descends, we must remember. 81 years ago tonight. Kristallnacht. https://t.co/u0zn13kOyW
Retweeted by Miles Brundage @togelius Only follow dog video accounts @earino @chrisalbon @Wohven_Tees I see all but one replyStates force the world into simplistic models to make it legible to bureaucracy, sometimes at the cost of great har… https://t.co/9M2Vdr44Du
Retweeted by Miles Brundage
11/9
2019
One of these headlines was generated by @OpenAI's neural network, GPT-2. Which one? Test your fake text detection… https://t.co/2G1dp7mRNg
Retweeted by Miles BrundageAlmost every expert I know correctly predicted that Iran was likely to expand its nuclear program if the US pulled… https://t.co/FC0DH8R5rK
Retweeted by Miles BrundageAn update: @geoffreyhinton has been in contact with Canadian immigration officials. They told him that anyone who… https://t.co/QKHnCHgCjY
Retweeted by Miles BrundageCalling all women in @datascience and #ML/#AI in #NYC! If you're new to @PyTorch and would like to learn more about… https://t.co/r77gdaVJyW
Retweeted by Miles BrundageThe cost of training a neural net on public cloud servers has fallen by 100x in two years. https://t.co/5SHgt8MoQB
Retweeted by Miles BrundageKelsey Piper in Vox on our paper on judgements about the badness of human extinction. @LuciusCaviola, @nsfaber https://t.co/fFfT8b4aMA
Retweeted by Miles Brundage
11/8
2019
New analysis (and graph with a mind-bogglingly large log scale) -- quantitative macrohistory of AI! We're uncerta… https://t.co/4lff7MQaKJ
Retweeted by Miles BrundageWe continue to work with our service provider to restore service, and we will provide updates when they become available.
Retweeted by Miles BrundageMore room opened up on the #WiML2019 waitlist. Add yourself NOW to be first in line for the next ticket drop 🔥🔥🔥… https://t.co/As9S04c39M
Retweeted by Miles BrundageReleased a new analysis showing that compute for landmark AI models from before 2012 grew at exactly Moore's Law.… https://t.co/ZBO61MEEAO
Retweeted by Miles BrundageWe've analyzed compute used in major AI results for the past decades and identified two eras in AI: 1) Prior to 20… https://t.co/3JVuW5qGi1
Retweeted by Miles BrundageA court in China sentenced to death a man accused of trafficking fentanyl to the United States after a joint invest… https://t.co/J19H3WV5fh
Retweeted by Miles Brundage @ArtirKel Replacement tweet :) https://t.co/683nWuepsqActually, New Yorker folks would call it coöperation.Disproportionately bothered that the official spelling of "cooperation" in OECD is as follows: Organization for Ec… https://t.co/NCvNuZadDbInterested in whether NLP models know anything about numbers? See our fun poster at 1:30. Spoilers: they do, BERT i… https://t.co/4wCeQ9nMf6
Retweeted by Miles BrundageFinally the largest model GPT-2 is out! Along with the code & model weights, they also published some interesting f… https://t.co/hC1JeGJKnB
Retweeted by Miles BrundageThe full GPT-2 model is out - but the conversation around responsible release norms needs to continue.… https://t.co/8ONggYlKnG
Retweeted by Miles BrundageNaturally, this is a single-authored paper: "Fast Transformer Decoding: One Write-Head is All You Need," Noam Shaz… https://t.co/rnAeVgF28e @lizjosullivan 🧐🧐🧐 @lizjosullivan Due to AI focus, but that is also partly because it’s newer and AI is hotter now, so don’t give em much foresight credit @lizjosullivan Enemy of the State was IMO one of the most forward looking surveillance themed films re: what forms… https://t.co/ukMf4w1Vx7So I'm excited to share this GPT-2 output detector by OpenAI build with RoBERTa, which we created a web interface f… https://t.co/Q5Km2wHgE6
Retweeted by Miles BrundageAt @huggingface, we believe the most efficient way to reduce the negative impact of human-level NLP is to educate a… https://t.co/RwRoWpUCyp
Retweeted by Miles BrundageWhat a cool idea https://t.co/I3znyRoeS7
Retweeted by Miles Brundage
11/7
2019
Self driving is the clearest example of unchecked large scale public experiments in machine learning but it's not a… https://t.co/egld0ebhPl
Retweeted by Miles BrundageBig models => good videos :) https://t.co/ddIdOB6Ov9
Retweeted by Miles BrundageTry the @OpenAI output detector model for yourself with this web demo: https://t.co/vBbyvgNd13 Enter a block of te… https://t.co/jAu17bVm6v
Retweeted by Miles BrundageOur team has released a GPT-2 style model for dialog: DialoGPT: Large-Scale Generative Pre-training for Conversat… https://t.co/xQ1i9ysAvl
Retweeted by Miles Brundage @ResObscura @RoyaPak @OpenAI Hey! We're also concerned about this and spoke to a journalist about it recently:… https://t.co/XRbnmniEugThank you to GPT-2 for helping me write this story 👍 https://t.co/pERxFFmXmS
Retweeted by Miles BrundageBlue-footed boobies dive bomb the water simultaneously https://t.co/IWWHLCk0nF
Retweeted by Miles Brundagedating in sf: sitting in a cafe nodding as some guy tells you “Peter Thiel backs outsiders... Jesus Christ was the ultimate outsider”
Retweeted by Miles BrundageI'm concerned we may start seeing a trend of folks implying bias is a solved problem. I heard 2 examples of this la… https://t.co/Apw8YlLp1Y
Retweeted by Miles BrundageExcited to share my summer intern project @facebookai “Emerging Cross-lingual Structure in Pretrained Language Mode… https://t.co/6TpvQl5bA1
Retweeted by Miles BrundageAwesome youtube channel focusing on interpreting AI and its meaning for society: https://t.co/n5zmtUX3Px by @JordanBHarrod
Retweeted by Miles BrundageWe’ll be presenting our poster at #emnlp2019 *this morning* on “Language Models as Knowledgebases”… https://t.co/VyzUbnFJby
Retweeted by Miles BrundagePlease check out the final @OpenAI report at the link below and please read @CTECMIIS portion where we fine-tuned t… https://t.co/M0RgKs4xxf
Retweeted by Miles Brundage @JohnWoodTorch @AdamDanielKing https://t.co/wlv3Tu8Sdl @JohnWoodTorch @AdamDanielKing You might be interested in the report we published earlier which has a section on this!
11/6
2019
At @aifdn, we are working hard to build detection tools along with our synthesis tools to protect abuse. Our first… https://t.co/fgsIe54wFu
Retweeted by Miles BrundageThe largest GPT-2 model is now LIVE! So happy to have participated in the detection work which marks my first (co-a… https://t.co/2VlZNRbhNb
Retweeted by Miles BrundageThank you @OpenAI @alecrad @WuTheFWasThat @jackclarkSF et al! We will make good usage of the full GPT-2 language mo… https://t.co/oikW4bb58F
Retweeted by Miles BrundageReally excited to be sharing out latest and final report on GPT-2, which is being published alongside the 1.5B mode… https://t.co/Y7VWmvSmXP
Retweeted by Miles BrundageThe 1.5 billion parameter GPT-2 (aka gpt2-xl) is up: ✅ in the transformers repo: https://t.co/KvUK5V7owl ✅ try it o… https://t.co/TprkUa4HEs
Retweeted by Miles BrundageWell here we are. Congratulations to all the hard working people at OpenAI for their diligent work and for taking t… https://t.co/wsVlyMAgw0
Retweeted by Miles BrundageP.S. Just wanted to make a final plug for checking out the report itself, which has tons of findings from OAI staff… https://t.co/fV7blDTricI guess our lunch break will be shorter than usual. Thanks @OpenAI ;) #gpt2
Retweeted by Miles BrundageThe "too dangerous to release" GPT-2 text generator is finally fully released! I've put it up on… https://t.co/IyXMlPu2Fy
Retweeted by Miles BrundageDetection model based on RoBERTa (via the @huggingface Transformers library) https://t.co/VKQ8kJfZGM
Retweeted by Miles Brundage @julien_c @huggingface @LysandreJik @jamieabrew We would be excited for you all to host it! We're currently focused… https://t.co/pyVKAnqUEAExtremely excited to share @CTECMIIS’s preliminary research on the threat posed by neural language models. We’ve be… https://t.co/0UiHToNbkj
Retweeted by Miles Brundage5. Finally, we’ve learned a ton both from those who agreed and those who disagreed with our actions, and everywhere… https://t.co/nuvOV9xEiw4. The AI community now has some new tools to use in similar situations (staged release, partnership-based sharing)… https://t.co/N7mIFkYOmT3. Tracking language model impacts is hard! We’ve done our best to assess the prospects for beneficial applications… https://t.co/mfuJWjvUiV2. A related point is the need for cross-organizational communication and collaboration. Working with others in the… https://t.co/n9unCyQ7WO1. One clear takeaway from the GPT-2 case is the critical importance of diverse disciplinary perspectives in addres… https://t.co/PXoph96ePVReally excited to be sharing out latest and final report on GPT-2, which is being published alongside the 1.5B mode… https://t.co/Y7VWmvSmXPExcited about the final model release of our GPT-2 staged release and all our research findings! This was a great t… https://t.co/7X8ZjCO5nX
Retweeted by Miles BrundageWe're releasing the 1.5billion parameter GPT-2 model as part of our staged release publication strategy. - GPT-2 ou… https://t.co/en3NeSXyNL
Retweeted by Miles Brundage @gwern Oh nice! Thanks!I assume the fine-tuning is pretty important, because I don't recall it coming up in the past and there seems to be… https://t.co/Eqn7q0iTWFCold take on the AlphaStar paper: I'm somewhat surprised that AlphaStar Supervised beat 84% of players, and wonder… https://t.co/kKyNNEkK2dLinking artificial and human neural representations of language “We find that language models with degraded syntac… https://t.co/nv34YMVFwI
Retweeted by Miles Brundage"Learning to Fix Build Errors with Graph2Diff Neural Networks," Tarlow et al.: https://t.co/8HRmRX0doB
11/5
2019
Interesting post from Google on a "celebrity [face] recognition" tool and associated safeguards: "“Celebrity” is c… https://t.co/yLdxawk2DJhe whomsoever that canst draw the motorola from this carriageway shall be kinge of all post brexit englandde https://t.co/Gr7FtEw78C
Retweeted by Miles BrundageOk, Boomer. https://t.co/GuIyv6gNBO
Retweeted by Miles Brundage
11/4
2019
Aight here's like, a whole bunch more pokemon. I've just started training a new one on a dataset of thumbnails from… https://t.co/4O6J4EAtxt
Retweeted by Miles BrundageIf at EMNLP, come talk to me at the poster for this paper at 16.30 on Wednesday, Hall 2A :) (paper with… https://t.co/AY9aUJZmmn
Retweeted by Miles BrundageHow do representations evolve as they go through the transformer? How does the Masked Language Model objective affe… https://t.co/8zBKTX68dK
Retweeted by Miles BrundageIt’s official - @POTUS has set the lowest refugee cap in history despite record numbers of refugees globally. https://t.co/CbwaB93Czq
Retweeted by Miles Brundage
11/3
2019
Applications open for doing a PhD at the UCL Centre for Doctoral Training in Foundational AI!… https://t.co/hYc9z5tKYh
Retweeted by Miles BrundageExcited (in my 1st tweet ever!) to announce Adversarial NLI: a new large-scale benchmark dataset for NLU, and a cha… https://t.co/PBibyyMYkS
Retweeted by Miles BrundageI found the financial inclusiveness + flexibility re: family structure stuff especially interesting."Halloween gives us a rare chance to be nice to our neighbors without any religious precedent. ... The actions unde… https://t.co/ecaD7QeCONWhen the system encourages most researchers to optimize for depth, there will be vast opportunities for the minorit… https://t.co/Ap0KwPSnGJ
Retweeted by Miles BrundageWe explore a simple approach to task-oriented dialog. A single neural network consumes conversation history and ext… https://t.co/nj8VV6uKmJ
Retweeted by Miles Brundage
11/1
2019

0