-
- Private work.
- Cuban still gonna run ops, but selling majority stake in mavs for 3.5B, bought in 2000 for 285M. Slightly >10% return per year.
- Sold to Adelson family, casino/gambling empire.
- Kramnik and Hikaru: https://www.chess.com/blog/CHESScom/regarding-recent-accusations
- Convector units serviced in apt.
- AWS is about 1/8th (~80B in 2022) of amazon’s revenue (~470B in 2022). Ecommerce is larger revenue. But AWS is ~75% (~19B in 2022) of total amazon operating profit (~25B of 2022).
- Claude is anthropic’s AI assistant.
- Uploaded photos/videos to icloud to get ~50GB back on device.
- Spotify 2023 wrapped.
Promise.race() to take a list of promises and return whenever the first promise returns, forwarding along the resolve/reject.- Side button still defaults to siri for ios. Can disable, but can’t customize to something like powerdown (as it used to).
- Played with bedrock a little bit.
- Connect your own private knowledge bases to help the FM train. Adds relevance/context to your genai app.
- Chat, perform actions, answer questions, search, generate images, summarize notes – APIs for everything.
- re:Invent.
- >50k in vegas, >300k online.
- 12th annual.
- Watched Selipsky’s keynote.
- New: S3 Express One Zone. Up to 10x faster than S3 Standard.
- New: Graviton 4. At least 30% faster than 3.
- Remember nitro’s virtualization is extremely efficient. Abstracts a lot so your available compute is fully allocated to your app.
- Guest: Jensen Huang, ceo/founder nvidia. Basically the new grace hoppers (cpu + gpu) are available to huge infra on aws.
- New: Trainium 2. New chip for ML.
- Guest: Dario Amodei, ceo/cofounder anthropic.
- Titan models are FMs from amazon.
- New: Bedrock Guardrails. Plug in your custom policies for responsible AI.
- Guest: Lidia Fonseca, cto pfizer.
- CodeWhisperer. Autocompletion, explain code, translate, etc. Basically copilot, but less generally-capable, and more tuned to aws.
- New: Amazon Q. AI assistant. Custom trained on all aws services and great for helping design. Respects data privacy, IAM, etc. Will integrate into the IDE soon, making it a nice merge with codewhisperer. The specific feature is called Code Transformation. Example datapoint: internal amazon team successfully upgraded 1000 applications from java 8 to 17 in 2 days.
- Guest: Matt Wood, tech vp aws. Deepdive into Q.
- Guest: Stephan Durach, svp bmw.
- New: Zero-ETL integrations from redshift to (1) aurora postgres (2) rds mysql (3) dynamo (4) opensearch.
- Ended with Kuiper – not sure why in aws keynote, probably cool factor of space. They’re going to not just offer public broadband – they’ll offer private enterprise networks as well.
- Elon and Sorkin discussing twitter, antisemitism, openai, more: https://www.nytimes.com/live/2023/11/29/business/dealbook-summit-news.
- Got spicy at times.
- Larry Page stopped talking to Elon after Elon brought Ilya to OpenAI.
- Still waiting for a reason to be published re Sam’s ousting. But sounds like Elon believes Ilya had good reason (unknown to us).
- NBA fouls.
- Did this before, worth doing again. The lower both numbers, the better (luckier, favorable treatment from refs, etc). Teams do NOT natively commit a substantial difference in fouling – it’s nonzero but not as large as the gaps below.
- 2022-2023 season, GS 29th in fouls (2nd most whistles blown against them) and GS 28th in opponent fouls (3rd least whistles blown against the opponent). For comparison (since they played last night): Kings 12th and 3rd.
- 2023-2024 season, GS 28th in fouls (3rd most whistles blown against them) and GS 14th in opponent fouls (avg number of whistles blown against the opponent). For comparison (since they played last night): Kings 24th and 8th.
- Supercontest.
- Checked the cache for a view key – yep, the value is just the rendered HTML (css and js still served from flask/gunicorn).
- Looks like the cache is max ~10MB (when everything is not expired)
redis-cli info memory.
- Elasticache serverless pricing.
- Storage: $0.139 / GB-hour, and I have .01GB = $1/mo.
- Requests: $0.0038 / million ECPUs, where an ECPU is basically 1KB of data read or written, and say I have 10 reads/writes every minute, each 100KB -> 1000 ECPUs/min -> $0.16.
- That would all be well and dandy, but the MINIMUM meter for a cache is 1GB. So my 10M cache doesn’t get the cost savings. With the minimum, storage becomes $100/mo. And requests are trivial after that.
- So no – probably not worth the price yet.
- Spreads.
- Should probably go back through and replace all “line” with “spread” – more common. Handicap is also used.
- This site shows (static) lines across multiple books: https://www.wagertalk.com/odds?sport=L1&type=1&date=2023-11-26&cb=0.6602684239755252
- The second number after spreads is (I believe) the moneyline for that spread. Usually around -110 for both teams to cover. Very close to 50/50, as spreads are designed to achieve. The lower the number, the more likely.
- SQS.
- Spent most of the day migrating the score fetch/commit to sqs/lambda/eventbridge.
- Punted the upgrade of picks to the same event/queue, once the pick interface is in react: https://gitlab.com/bmahlstedt/supercontest/-/issues/232.
- Created the FIFO queue,
supercontest-scores.
- Created the 2 new lambdas, the fetcher (reads from api, writes to queue) and the committer (reads from queue, writes to db).
- In this process, plugged a few things in:
- Auth: Lambda <-> SQS
- Auth: Lambda <-> RDS
- Trigger: EventBridge <-> Lambda
- Trigger: SQS <-> Lambda
- Boto3 in the lambda function to write to sqs, simple.
- Remember you have to pass MessageGroupID with FIFO queues. It’s for logical separation, guaranteeing only-once within single msg groups.
- Sent the same msg a few times from lambda test. SQS properly deduplicated (there was only one message in the queue). But there’s a piece of metadata on the msg called
receive_count which ticks up for every duplication.
- Remember that a HUGE benefit of queues is the abstraction of producers and consumers. If you have one trigger and one response, one writer and one reader – a queue loses much of its value. Just directly plug the 2 services.
- It’s still nice to atomize the various pieces, and queues make you think about the producer/consumer split still.
- Submitted picks.
