_009

Back in 2005, I remember working on startups running on Scrum principles. It worked well at the time, we where able to ship, grow the team, and move forward with a nice few-features-per-week cadence, working remotely, on a small team; less than 10. Tt always worked fine, but very centralized and slow, as all-things-dev were at the time.

I worked with ActiveColab in 2007, Skype 2007, Yammer 2009, Trello 2011, Pivotal Tracker 2013, Trello 2016, Confluence 2022, Slack 2013, Google Meet, and sometimes I think, scrum became _less-relevant_ over the years as more advanced product management tools became the norm and the product manager role matured by leveraging them.

These days, it's not rare to see lead developers manage kanban-like boards very effectively, releasing on time, with grace, without the need of a scrum master to coordinate efforts.

I do like asynchronous scrum daily standups using http://geekbot.com on slack, when on-site or/and distributed and doing sprints. I seen this work well on startups going from pre-seed to series B.

Personally, I am fascinated with team dynamics and how they've changed over the years. We are definitely living the best of times as a developer and I still see sparkles of well-applied scrum every now and then that works nicely.

For me, one of the coolest thing about the Robocop movie series is that, each Robocop movie is about the next Robocop. Robocop 1 is about Murphy being the first Robocop, Robocop 2 is about Cain becoming the second Robocop, and Robocop 3 is about the asian looking Robocops -- the third generation.

> my friends call me murphy you call me robocop

AWS is the next Oracle

AI doesn't scale well. Problems get worst as you make your model bigger and more generalized. To make things worst, data, model architecture, precision, hardware, affect your model performance in ways that are hard or impossible to anticipate.

If you watch Tesla's AI presentation, https://www.youtube.com/watch?v=HUP6Z5voiS8, you will notice that they have multiple AI's stacked on each other, which IMO is a step back from truly e2e multimodal AI system. So even with their custom fancy hardware, multimodal is too hard.

I wonder, wouldn't it be better to use geo fencing (using H3), and have the car download the model depending on the zone where it is driving? And optimize multiple models based on "driver engagements"? This could fix the problem of zones where there are particularities in the driving, road, or human activities, and allow for model optimization to happen on a smaller vector space than the whole world. For example, why not have a model for US highways, LA, New Deli, UK, so on.

Tesla also knows where the cars are, and control their expansion plans worldwide, which could inform model prioritization roadmap.

In my mind, it will be easier to test, debug, label, optimize, and guarantee quality to users, that at the end of the day, without knowing exact statistics, I am dare to say spend more than 70% of the time driving around the same county/city/area/town?

One of the most brilliant engineers out there. A true madman with an old hacker mentality that is nowhere to be seen these days, except for maybe George Hotz...

Old days where different, today, it's about leetcode and being overly happy on zoom calls, and playing along investors playbooks... Capitalist only left the hoodies, and that's because another 100m funded startup from their portfolio are selling them.

Feeling nostalgic...

While garbage in, garbage out may seem like a bad policy to the user, to the AI system, it means that it can have a closed feedback loop, where the final code (the solution) can be linked to the initial input, regardless if the input was garbage or not.

I would say that anything that can be stated as a large-scale supervised reinforced learning problem is a gold mine -- if the output of course, has value and supervision is free.

Tesla self-driving and Comma.ai, from an eagle's eye view, exploit the same concept.

Fascinating stuff. There nice trajectory animations on wikipedia, https://en.wikipedia.org/wiki/OSIRIS-REx

From wikipedia:

> Such asteroids are considered "primitive", having undergone little geological change from their time of formation. In particular, Bennu was selected because of the availability of pristine carbonaceous material, a key element in organic molecules necessary for life as well as representative of matter from before the formation of Earth. Organic molecules, such as amino acids, have previously been found in meteorite and comet samples, indicating that some ingredients necessary for life can be naturally synthesized in outer space.

Great!

Wow, hardware already made it to low-orbit, https://upsat.gr/?p=418 (2017-04-20)

> At April 18th 11:11 EDT at Cape Canaveral in Florida, an Atlas-V rocket launched a Cygnus cargo spacecraft to dock to the Internation Space Station with supplies and scientific experiments. Among its cargo UPSat, the first open source hardware and software satellite bound to be released in orbit by the NanoRacks deployment system on-board ISS in the coming weeks.

Apart from search, ANNs can be use for recommendations, classification, and other information retrieval problems.

Currently, ES and Solr, both based on Lucene, can't really manage vector representations, as they are mainly based on inverted indexes to n-grams.

ANNs potential applications extend to audio, bioinformatics, video, among any modality that can be represented as a vector. All you need is an encoder! How nice.

Faiss is definitely powerful. I have been running experiments using 80 million vectors that map to legal documents, and vectorizing protein-folds (using Alphafold). While it is an interesting technology, at this moment, perhaps for my usecases, I see it more as a lib or tool than a full-featured product like ES or Solr.

For instance, ATM, updating a Faiss index is a non trivial process, with many of the workflow tools you would expect in ES missing. There is also the problem of encoding the input into vectors, which takes a few milliseconds (do you batch, parallelize, are you ok with eventual consistency?).

I recently been found with pgvector (postgres + vector support) https://github.com/ankane/pgvector. Perhaps less performant, but easier to work with for teams. With support of migrations, ORM, sharing, and all the postgres goodies.

Another interesting/product-ready alternative is https://jina.ai.

And Google's ScaNN, https://www.youtube.com/watch?v=0SvrDtnUgV4

I remember meeting with KimDotCom lawyer a few years ago in SF. A very flamboyant-person. One of those guys who makes a big impression on you from the first sight.

While KimDotCom is an impressive hacker, I am not sure that I would love to be the target of the FEDs around the world as he is still in the eye of the US.

Fun story about MegaUpload.

I once lived in Argentina back in the MegaUpload days. At the time, piracy was the norm (not only in Argentina), the gov didn't care, and people where selling pirated, burned DVD on the streets. This was a downtown, a high transited area.

Then MegaUpload started to grow like fire, and I remember that starting at 4 pm, the internet would get awfully slow. As people get off their jobs to download the latest movie or episode out there. Then PopcornTime, and things got even worst. Cant find the stat, but I remember something along the line of 60% of Buenos Aires traffic being MegaUpload's at peak time (4-10pm), which caused a lot of controversy at the time.

Old days...

If you are looking to do semantic search (Cosine similarity) + filtering (SQL) on data that can be represented as vectors (audio, text, video, bio) I suggest, https://github.com/ankane/pgvector