G
Google AI Blog
ai.googleblog.com · Company blog · on HN since 2018
Google's research team posts on machine learning, speech, vision, and language models, plus occasional general engineering pieces.
What HN saysHN commenters cite it as a source for Google's AI research, such as the ReAct model post, when arguing about where AI is heading.
596HN people shared it
329linked it in comments
866stories submitted
16Kstory points
6 months agolast shared
2018people per year2026
Most-shared posts
- 1Nearly all binary searches and mergesorts are broken (2006)30 people · 173 points · first shared 2018 · 195 comments on HN
- 2Federated Learning: Collaborative ML Without Centralized Training Data18 people · 10 points · first shared 2018 · 1 comments on HN
- 3Google Duplex: An AI System for Accomplishing Real World Tasks Over the Phone12 people · 1.9K points · first shared 2018 · 750 comments on HN
- 4Lyra: A New Very Low-Bitrate Codec for Speech Compression11 people · 70 points · first shared 2021 · 25 comments on HN
- 5An All-Neural On-Device Speech Recognizer11 people · 13 points · first shared 2019 · discussion on HN
- 6Google Research, 2022 and Beyond: Language, Vision and Generative Models11 people · 12 points · first shared 2023 · discussion on HN
- 7A Multi-Axis Approach for Vision Transformer and MLP Models11 people · 6 points · first shared 2018 · 1 comments on HN
- 8Experimenting with Automatic Video Creation from a Web Page10 people · 20 points · first shared 2020 · discussion on HN
- 9Pathways Language Model (PaLM): Scaling to 540B parameters9 people · 235 points · first shared 2022 · 207 comments on HN
- 10Language Models Perform Reasoning via Chain of Thought8 people · 3 points · first shared 2022 · discussion on HN
- 11Quantum Supremacy Using a Programmable Superconducting Processor7 people · 701 points · first shared 2019 · 179 comments on HN
- 12Turbo, an Improved Rainbow Colormap for Visualization7 people · 494 points · first shared 2019 · 121 comments on HN
- 13Reformer, the Efficient Transformer7 people · 175 points · first shared 2020 · 43 comments on HN
- 14Dopamine framework: Fast prototyping of reinforcement learning algorithms (2018)7 people · 106 points · first shared 2018 · 15 comments on HN
- 15Super-resolution on the Pixel 37 people · 77 points · first shared 2018 · 31 comments on HN
What HN people wrote when they linked it
Manifretto · 2021-02-23
I read the article and i'm not following the argument delivered at all.
There is no real proof to this.
I'm following and reading research @ google (stuff like this https://ai.googleblog.com/ and other sources) for ages now and NOTHING indicates an 'implosion'.
It is strong research with real and constant results.
I have no idea why the autor would even consider using the word 'implode'.
Its not rocket science that data is biased and it just will continue be researched and a solution will be found. For the single reason that biased systems in certain areas will not deliver the results you need to use it properly.
AllanHoustonSt · 2020-03-26
Unfortunately I agree with you Jane Street's tech blog and published lecture series are great and probably the best I know of. They make for great advertising to boot. Lecture series like Jane Street's are common in larger trading firms (more than 100-150 headcount) but I haven't seen one of JS' caliber yet.
If you welcome general corporate blogs, I think Google AI's technical blog is quite good. Their frequent publications on distributed computing strike a good balance of academic and practical ideas:
https://ai.googleblog.com
Cloudflare also has one of my favorite more engineering focused blogs.
deadmutex · 2023-05-22
If you don't want to have to worry about charging things, or wearing them to bed, Google Nest Hub is great: https://store.google.com/us/product/nest_hub_sleep_sensing?hl=en-US .
It can also track sleep stages, etc. This has been on sale for < $75, which is way less than a wearable, etc.
https://ai.googleblog.com/2021/03/contactless-sleep-sensing-in-nest-hu…
https://ai.googleblog.com/2021/11/enhanced-sleep-sensing-in-nest-hub.h…
scrollaway · 2023-04-06
I don't remember the context, but I have seen properly-prompted GPT-4 proactively ask questions.
It's also worth noting that the future is multi-layered. The Reason+Act model (https://ai.googleblog.com/2022/11/react-synergizing-reasoning-and-acti… should be excellent at getting the LLM to analyse its own output and inquire about missing pieces of knowledge.
nopinsight · 2023-03-25
This is a very interesting hypothesis that could be quite true for living beings. What I disagree with is that having an animal-like body is necessary for the process of forming a world model. A simulation could be sufficient. And there is already work on that front. (Also, I would not characterize deep-learning-based AI as trying to form propositional knowledge. In fact, its great performance partly stems from not dealing with propositional knowledge directly.)
If a body is in fact necessary, PaLM-E could be paving a way toward it as well.
https://ai.googleblog.com/2023/03/palm-e-embodied-multimodal-language.…
vladf · 2023-03-17
This was done last year: https://ai.googleblog.com/2022/02/can-robots-follow-instructions-for-n…
ux-app · 2023-02-20
>However it is still just pattern matching
the visual cortex in your brain is also "just a pattern matching" system. guess it's not very impressive by your standard.
This[1] isn't my example (it's from another HN user), but if you work as a programmer and you're not absolutely jaw on the floor astonished by this example then I don't know what to say.
Explaining[2] the emergent behaviour is literally cutting edge research. Hand waving this behaviour away as just "probabilistically generating a likely next word" is ignorant.
It's amazing in similar ways to Conway's Game of Life.
[1] https://imgur.com/HOEnxYb
[2]https://ai.googleblog.com/2022/11/characterizing-emergent-phenomena-in…
achow · 2023-01-20
Quite possible (ChatGPT related).
Considering this was the Tweet that he sent yesterday.. https://twitter.com/sundarpichai/status/1615820298305118221
As a followup to AI blogpost (which is clearly in response to ChatGPT) https://ai.googleblog.com/2023/01/google-research-2022-beyond-language…
kache_ · 2022-12-22
>There's still no AI that can fold laundry
We're actually really close to general robot agents that operate in your home. Check out googleAI's saycan & RT-1 systems
https://ai.googleblog.com/2022/12/rt-1-robotics-transformer-for-real.h…
nl · 2022-12-17
This is untrue.
Minerva in particular does step by step reasoning:
> we present Minerva, a language model capable of solving mathematical and scientific questions using step-by-step reasoning
https://ai.googleblog.com/2022/06/minerva-solving-quantitative-reasoni…
They use a technique called chain-of-thought to make it reason problems out step by step.
As Norvig shows, this reasoning may sometimes be wrong, but that's a different issue.
It's worth repeating this because apparently it's not widely known: these models are following a step-by-step process we'd recognise as reasoning in humans.
lpage · 2022-12-07
Joshua Bloch did a great post [1] on how nearly all standard library binary searches and mergesorts were broken (as of 2006) due to this exact issue. The punchline is that the bugs started cropping up when the need and capability to sort O(2^32) element arrays arose.
[1]: https://ai.googleblog.com/2006/06/extra-extra-read-all-about-it-nearly…
derangedHorse · 2022-12-04
OpenAI has been making the most noise online because of how open they've made their recent chatbot, but Google has been on this for a while. Earlier this year they had a blog post [1] about LaMDA which doesn't seem too far off in capability from OpenAI's projects. They've also made a lot of other strides in their research [2] that kind of goes under the radar because they haven't been synthesized into products yet (at least not in the ways we'd expect them to).
[1] https://ai.googleblog.com/2022/01/lamda-towards-safe-grounded-and-high…
[2] https://ai.googleblog.com/
andrewia · 2022-11-01
I think it's processing power and engineering effort. I got a Sony RX100M2 for my mom and it has the same computational photography techniques that Google released the following year (https://ai.googleblog.com/2014/10/hdr-low-light-and-high-dynamic-range… But Sony's image stacking is only in "Superior Auto" mode, and is only used when necessary. Google's implementation does a lot of advanced work, including selecting and blending parts of the photo depending on motion, that Sony doesn't do. I assume Sony's imaging engineers have less expertise in advanced processing, and didn't have the resources to implement the features that Google did. Sony also has to devote engineering resources to other features - a lot of their sales are to photographers that will edit in post (RAW), and later, videographers. So features that are only in "auto" mode may have limited budget.
imranq · 2022-10-28
Stanford grad courses are excellent. Just following the exercises will get you very far:
* AI Foundations: https://stanford-cs221.github.io/autumn2022/
* Meta Learning: https://cs330.stanford.edu/
* Theoretical ML: https://cs229.stanford.edu/
* NLP: https://web.stanford.edu/class/cs224n/
* Computer vision: http://cs231n.stanford.edu/
* Reinforcement Learning: https://web.stanford.edu/class/cs234/
Read Blogs:
* DeepMind: https://www.deepmind.com/blog
* Google Research: https://ai.googleblog.com/
Replicate papers that you are interested in with a twist specific to your use cases
* RL Spinning UP:https://spinningup.openai.com/en/latest/
pdw · 2022-08-15
That's a great book. Its binary search is buggy though :)
https://ai.googleblog.com/2006/06/extra-extra-read-all-about-it-nearly…
gobengo · 2022-06-17
> Can these computers actually do something better than their classical counterparts, right now?
The answer is yes. Sometimes this is referred to as 'quantum supremacy' and some researchers at Google declared it in 2019.
https://www.nature.com/articles/s41586-019-1666-5
https://ai.googleblog.com/2019/10/quantum-supremacy-using-programmable…
aidenn0 · 2022-05-23
This is quite suspicious considering that google AI research has an official blog[1], and this is not mentioned at all there. It seems quite possible that this is an elaborate prank.
1: https://ai.googleblog.com/
blihp · 2022-04-20
> The chip design itself should be the secret sauce. Not the tools you make the chip with.
The secret sauce is generally whatever gives one a competitive advantage. Businesses typically open things up when the want to reduce the cost of something and/or cause pain for someone else (i.e. killing their cash cow), not because they're benevolent and want to share.
> I challenge you to answer your own question in reverse: are any companies other than Nvidia embarking up AI/ML chipmaking in a closed fashion? There probably are, let's follow & watch them.
Isn't Google's floorplanning work closed source? https://ai.googleblog.com/2020/04/chip-design-with-deep-reinforcement.…
sdenton4 · 2022-03-06
There's ample literature showing that neutral vocoders generalize to new speakers not seen in training. IME, once you train on around ten speakers the models generalize quite well to new speakers, and in practice training sets for production models are often using hundreds of speakers for training.
There are also non autoregressive GAN vocoders which work great, with a tiny fraction of the compute of the original WaveNet. Soundstream is a good example of this.
My own work on bird song separation uses non AR models for generating separation masks. It works great and generalizes to new habitats and species without difficulty. https://ai.googleblog.com/2022/01/separating-birdsong-in-wild-for.html…
Calling architecture variants a hack is risking no true Scotsman arguments... The only true neutral networks are the ones that don't work, perhaps?
vishnu_ks · 2022-01-01
A couple of my favorite technical blogs
- https://openai.com/blog/
- https://ai.googleblog.com
- http://antirez.com/
- https://research.facebook.com/blog/
- https://martin.kleppmann.com/archive.html
- https://devblogs.microsoft.com/oldnewthing
- https://www.allthingsdistributed.com/
- https://jvns.ca/
- https://www.johndcook.com/blog
- http://feross.org/
- https://ridiculousfish.com/blog/
- https://blog.andrewcantino.com/
- https://nickcraver.com/blog/
- https://blog.codinghorror.com
I have been also building an engineering blog aggregator called https://diff.blog to make it super easy to disover and follow blogs. Would love if you folks can give it a try :)
jacobparker · 2021-10-03
I highly recommend SF. The trick is you must use the books via coqIDE or something equivalent; viewing the rendered HTML is almost pointless at best, but probably counter-productive.
Binary search would be a good exercise. I think SF volume 1 would be more than enough, with the exception of deciding how to model arrays (you'd probably want to look up Coq's persistent arrays/PArray system for this.) SF volume 3 does binary search _trees_ which is similar but avoids the array problem (as is customary in functional languages.)
However, Coq (by default) also uses infinite precision numbers. This works very well for proofs but isn't necessarily realistic. This is where some extra fun comes in: fixed precision arithmetic is a very common source of bugs for binary search: https://ai.googleblog.com/2006/06/extra-extra-read-all-about-it-nearly…
murphyslab · 2021-08-29
The writing is great, but I keep wondering: Why do so many physicists use Jet or other non-perceptually uniform rainbow-like colourmaps for their graphs? Shouldn't that be part of great science communication, given that about 1 in 12 men have some form of colour blindness?
There are so many other visually appealing options [0] that are attractive while still being perceptually uniform. There's even Google's Turbo [1] colourmap for those who refuse to give up rainbow-like colourmaps.
[0]: https://matplotlib.org/stable/tutorials/colors/colormaps.html
[1]: https://ai.googleblog.com/2019/08/turbo-improved-rainbow-colormap-for.…
mchusma · 2021-08-15
Googles duo already does this. In my experience, it's amazing.
http://ai.googleblog.com/2020/04/improving-audio-quality-in-duo-with.h…
cgreerrun · 2021-06-24
This site keeps track of them:
https://github.com/erikbern/ann-benchmarks#glove-100-angular
ScaNN is currently SOTA:
https://ai.googleblog.com/2020/07/announcing-scann-efficient-vector.ht…
scoopertrooper · 2021-05-18
There have been great strides made in ai compression in recent years. I wonder if they are going to leverage that.
https://blogs.nvidia.com/blog/2020/10/05/gan-video-conferencing-maxine…
https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.…
https://aibots.my/officialBlog/deepminds-ai-agent-muzero-could-turboch…
logicslave · 2021-05-13
Read this blog:
https://ai.googleblog.com/
It maybe somewhat removed for what youre looking for, but heres the secret:
You cant speak intelligently and usefully about AI unless you have some kind of understanding about the cutting edge and whats possible. So read the newest research and connect it with your business
harias · 2021-04-16
Sounds a lot like Google's GAN for fantastical creatures[0]. Labels to photorealism seems to be the core idea behind both
[0] https://ai.googleblog.com/2020/11/using-gans-to-create-fantastical.htm…
ml-thoughts2018 · 2021-04-09
Annoy and ScaNN are two good options - ScaNN is too new to have solved your previous problem, but Annoy would have been great.
- https://github.com/spotify/annoy
- https://ai.googleblog.com/2020/07/announcing-scann-efficient-vector.ht…
mfkp · 2021-03-02
After reading vidarh's comment below ("The biggest challenge with evaluating all of these, is that once you've listened to a comprehensible version of one of these samples, they all sound more intelligible."), I recommend trying to listen to the Lyra version first, then listen to the original next. It actually does sound very strange when you don't have the true audio parsed in your head already. I'm not sure if I would like having to do the extra work to translate this in my head on every conversation.
It's more apparent in the video example from the Google post, I won't spoil it but there's a word that starts with B that sounds very funny in the Lyra version (listen to that one first): https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.…
walrus01 · 2021-03-01
the 3kbps example here with the 'bread with chocolate filling inside' video is frankly amazing, how good it is compared to the original.
https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.…
It is unfortunate that for now this appears to be proprietary, closed source and being treated as a google competitive advantage over others, unlike opus which is fully open.