If you have a field like and acl_id or some other context information on the data that is linked closely to a user's files. You can pass in the user's set of those field values to the vector database to pre-filter the results and do a permissions post check with a fairly relevant set.

The vector db definitely has to do some heavy lifting intersecting the say acl_id normal index with the nearest neighbors search but they do support it.

Does anyone have data on how much better these 1M token context models produce better results than the more limited windows alongside certain RAG implementations? Or how much better in the face of RAG the 200k vs 1M token models perform on a benchmark?

The founding fathers didn't see the corrupting influence of political parties on the separation of powers. It's covered quite well in this SMBC related comic: https://www.lawsandsausagescomic.com/comic/101

While I'm sure some of these predictions will come true and we'll have to layer some additional prompting and context engineering in here...

MCP seems to be the ultimate inheritor of ReSTful architectural practices as described by Roy fielding but instead of needing a human on the other end to interact with links (hateos) you can have an LLM bridging the gap between user intent and concrete subsequent calls.

So many issues around versioning and the fragility here will likely start to disappear. Like embeddings I'd put a bet on this being one of the major unexpected architectural improvements that lets us build our scaffolding a bit higher before it collapses to use a Steve Jobs analogy, LLM's at key glue points to create looser coupling.

The problem with this line of thought is that everyone just uses it to hate on people they already hated on. You go after rich tech monopolies, progressives generally go after big business, conservatives use it to go after the biggest business of all, the government.

We all just need to focus on monopoly in all its forms, instead of letting the politicians continue to separate us using division without difference.

And they managed to get it flagged. The longer HN mods pretend this is just natural flagging because folks are sick of the same old topics and not a coordinated effort to control the narrative, the more I'm going to start seeking alternative sources for interesting news.

At the very least someone showing data that in aggregate there are just more follow-on duped stories about things and they're letting one through un-flagged (ideally the top up voted one) to show that there is or isn't bias creeping in via the flagging system would be helpful in re-establishing trust.

I probably have a unique view as I view HN through an RSS feed of posts with over 100 up votes. Every single time I see a post critical of X or Musk and click through the story has been flagged. I'll try to do data analysis via that lense and see what it turns up.

As someone that criticized a number of their employers API's for not being sufficiently ReSTful especially with regards to HatEoS, I eventually realized the challenge is the clients. App developers and client developers mostly just want to deal with structured objects that they've built fixed function UX around (including the top level) and desire constructing URLs on the client. It takes a special kind of developer to desire building special mini-browsers everywhere that would require hateos and from the server side.

I think LLM's are going to be the biggest shift in terms of actually driving more truly ReSTful APIs, though LLM's are probably equally happy to take ReST-ish responses, they are able to effectively deal with arbitrary self describing payloads.

MCP at it's core seems to design around the fact that you've got an initial request to get the schema and then the payload, which works great for a lot of our not-quite-ReST API's but you could see over time just doing away with the extra ceremony and doing it all in one request and effectively moving back in the direction of true ReST.

I find having a therapist for an hour each week and then using chatgpt or gemini for specific scenarios that pop up providing some context from my understanding in therapy as well as telling it to be brutally honest with me and don't mirror or trying to make me feel better leads to some pretty useful insights into my psyche off cycle from therapy.

Please note this is after at least a decade of therapy and couples therapy so I've got a solid base of self insight that I'm working from.

yeah I've been thinking about them as stochastic content addressable memory. You can put as many next = userInput; while(true's) { next = mem[next]; } around them as you need in different forms. Single shot. Agents. etc and get wildly cool results out, but it's gated by some of the limitations there.

Over a decade ago I was starting a java process up with local MySQL and getting 45k rps read, request per thread with load wasn't hard to achieve. Not sure why this is an accomplishment.