[Acquired] Google: the AI company (Part 1)


You can't say you understand today's AI landscape without listening to this massive (4 hours!) Acquired episode on Google, focusing on its AI roots. Over three episodes, Acquired has a little over 12 hours worth of podcast just on Google! Well worth it IMO for the greatest business in history.

Selected highlights:

[07:23] basically every single person of note in AI worked at Google with the one exception of Yann Le Cun who worked at Facebook

This is truly mind-bending to think about, especially considering that Google is (at the moment) not the first name that comes to mind when we think about AI (LLMs) today. But the real kicker comes a few minutes in when we learn that did you mean? (launched in 2001!!) and google translate (2006) are the first practical application of language models to its search business which made it exponentially more effective. About 25 years ago, Google was already running machine learning in production, at fantastic scale (about 15% of Google’s entire data center infrastructure).
In 2003 AdSense uses PHIL to "understand" the advertiser page and run the Google Ad corpus against them. Again, another massive advantage for Google core business and again, 20 years before the collective AI hype.

[17:22] Maybe 10x engineers exist? If they do, Jeff Dean has to be at least 1000x: Jeff’s work with the team gets that average sentence translation time down from 12 hours to 100 milliseconds. Then they ship it in Google Translate and it’s amazing. A 432000x improvement!
Jeff is so good that they start making up Chuck Norris facts about him. My favourite: Jeff Dean's PIN is the last 4 digits of pi 😂

[35:10] Cat paper, which allows Google to build a deep neural network that can classify video content on Youtube, ultimately leading to a recommendation system working on the actual content and not on author-supplied information like title, description or tags.
Youtube is no longer a destination site.

The same technology will also be used for copyright infringement and content filtering. In 2012! Ten years before ChatGPT.
At this point it was all still running on CPUs, but it was all about to change when AlexNet was designed to run NVIDIA GeForce GTX 580s.

[1:25:00] In 2016 Google uses DeepMind to achieve a 40% reduction in the energy required to cool data centers! Again, AI has a massive, very real impact on Google bottom line.

[1:40:22] In Spring of 2014 Google buys 130 million worth of GPUs from Nvidia. This one single order is 3.25% of Nvidia revenue at the time. And with this new fast hardware, Google starts adding AI to Google Photos, Gmail autosuggestion, AdWords predicting interests, speech recognition on Android but at the same time Google realizes that they are just going to be shipping hundreds of millions soon to be billions of dollars over to NVIDIA every year for the foreseeable future.
So they set to work on the TPU and the TPU was designed, verified, built, and deployed into data centers in 15 months. FIFTEEN MONTHS!!!

And a great trivia about this rollout: they fit the TPU into the form factor of a hard drive, so it could actually slot into the existing server racks. You just pop out a hard drive and you pop in a TPU without needing to do any physical re-architecture. Amazing and clever!

And just to give an idea of the absolute scale of the TPU deployment (emphasis mine):
today it’s estimated Google estimated has 2–3 million TPUs. For reference NVIDIA shipped—people don’t know for sure—somewhere around 4 million GPUs last year.

People talk about AI chips like it’s this one horse race with NVIDIA. Google has an almost NVIDIA-scale internal thing making their own chips at this point for their own and for Google Cloud customers. The TPU is a giant deal in AI in a way that I think a lot of people don’t realize.

To be continued. 

Read my other AI-related posts.

Popular posts

Opengrep quickstart

Since I could not find a quickstart to run opengrep with the full set of rules from their fork I thought I'd document what I found out. Setup Download the opengrep binary from github and make it executable with chmod +x . Clone the rules repo: git clone git@github.com:opengrep/opengrep-rules.git and clean it up to make it usable to opengrep: cd opengrep-rules rm -rf ".git",".github",".pre-commit-config.yaml", "elixir", "apex" find . -type f -not -iname "*.yaml" -delete rm -rf .github rm -rf .pre-commit-config.yaml Ensure opengrep can load the rules with: opengrep_manylinux_x86 validate . The same can be done for custom rules maintained in a separate repository. AFAIU Multiple repositories can be specified by repeating -f option as needed, see below. We are now ready to scan a repo, from the repo root directory run: opengrep_manylinux_x86 scan \   -f <path_to>/opengrep-rules \   --error \   --exclude-rule=VAL some ti...
Read more

Mirth: recover space when mirthdb grows out of control

I was recently asked to recover a mirth instance whose embedded database had grown to fill all available space so this is just a note-to-self kind of post. Btw: the recovery, depending on db size and disk speed, is going to take long. The problem A 1.8 Mirth Connect instance was started, then forgotten (well neglected, actually). The user also forgot to setup pruning so the messages filled the embedded Derby database until it grew to fill all the available space on the disk. The SO is linux. The solution First of all: free some disk space so that the database can be started in embedded mode from the cli. You can also copy the whole mirth install to another server if you cannot free space. Depending on db size you will need a corresponding amount of space: in my case a 5GB db required around 2GB to start, process logs and then store the temp files during shrinking. Then open a shell as the user that mirth runs as (you're not running it as root, are you?) and cd in...
Read more

From 0 to ZFS replication in 5m with syncoid

The ZFS filesystem has many features that once you try them you can never go back. One of the lesser known is probably the support for replicating a zfs filesystem by sending the changes over the network with zfs send/receive. Technically the filesystem changes don't even need to be sent over a network: you could as well dump them on a removable disk, then receive  from the same removable disk.
Read more