Note that GPT-5.6-Sol is ~2x as token efficient as Opus/Fable, so a 258K max is equivalent to ~516K on Claude models. You also may just be fooling yourself, as context rot is still a problem[1][2] and there's evidence compaction can be equivalent or better than longer context windows for agentic tasks[3]. Obviously it would be best if the models could reason over 1M context like they do over 256K context, but they can't yet.
it does not require training a model from scratch for it to be updated. the entire LLM training process is iterative. essentially each step (there are hundreds/thousands for a training run) yields a complete, usable model. tokens with updated data can be added on top essentially at any time in the future.
You are correct. As a point of comparison: almost ten years ago at Segment we had a single Aurora PostgreSQL instance with ~50T of data, it was used to index potential identity data in a much larger corpus of files stored in S3.
Biggest thing to watch out with this approach is that you will inevitably have some failure or bug that will 10x, 100x, or 1000x the rate of dead messages and that will overload your DLQ database. You need a circuit breaker or rate limit on it.
This is explicitly called out in the blog post in the trade-offs section.
I was one of the engineers who helped make the decisions around this migration. There is no one size fits all. We believed in that thinking originally, but after observing how things played out, decided to make different trade-offs.
> It's a much better answer to hook up everything on Ethernet that you possibly can than it is to follow the more traveled route of more channels and more congestion with mesh Wi-Fi.
Certainly this is the brute-force way to do it and can work if you can run enough UTP everywhere. As a counterexample, I went all-in on WiFi and have 5 access points with dedicated backhauls. This is in SF too, so neighbors are right up against us. I have ~60 devices on the WiFi and have no issues, with fast roaming handoff, low jitter, and ~500Mbit up/down. I built this on UniFi, but I suspect Eero PoE gear could get you pretty close too, given how well even their mesh backhaul gear performs.
This seems pretty bad from the headline but there's no evidence of any in-the-wild exploits or if there is a feasible real-world exploit here. Some other domino(s) have to fall before it allows RCE. For instance, browser-based exploits are blocked by SELinux restrictions on dlopen from the downloads path.
Magnetic hard drives are 100X cheaper per GB than when S3 launched, and are about 3X cheaper than in 2016 when the price last dropped. Magnetic prices have actually ticked up recently due to supply chain issues, but HAMR is expected to cause a significant drop (50-75%/GB) in magnetic storage prices as it rolls out in next few years. SSDs are ~$120/T and magnetic drives are ~$18/T. This hasn't changed much in the last 2 years.
Is it bullshitting to perform nearly perfect language to language translation or to generate photorealistic depictions from text quite reliably? or to reliably perform named entity extraction or any of the other millions of real-world tasks LLMs already perform quite well?
> For e-commerce workloads, the performance benefit of write-back mode isn’t worth the data integrity risk. Our customers depend on transactional consistency, and write-through mode ensures every write operation is safely committed to our replicated Ceph storage before the application considers it complete.
Unless the writer is always overwriting entire files at once blindly (doesn't read-then-write), consistency requires consistency reads AND writes. Even then, potential ordering issues creep in. It would be really interesting to hear how they deal with it.
none of the things you cited are “miracle weight loss drugs.” they are things people did to lose weight. these are the first class of drugs that actually cause people to lose weight.
No, wide events are quasi-relational. Read the literature please. I’ve never seen a structured log implementation that is anything remotely relational, it’s just a bunch of random junk thrown into a pile with maybe some common tags.
architecture docs are scant, but it appears this horizontally scales channels by adding "broker" instances and horizontally scales consumers by adding "worker" instances? i.e. a broker only needs to ever broadcast a message to the maximum number of worker instances in the cluster.