elij

I just spawn a subagent in the cheapest range (for example flash-lite) to summarise a tool use. It's the only way that has worked based on my benchmarks and generalises well.

This is actually quite easy to implement at the harness level and the NER can be way more naive because of the typical nature of LLM dialogue (programming, long running tasks etc).

I'm using the 30b MOE model on same spec with 65k tokens as a sub agent with tooling and it absolutely writes decent code. The dense 9b I agree wasn't great.

if you're a linux person consider a routing on host setup with FRR with /32s. As every host is a /32 network you can focus more on the aspects of BGP rather than TCP/IP.

I asked Bard for the best Arsenal team and it was able to choose current players. It was also able to say how many goals they scored in current season as of today.

same issue with networking (specifically per packet heuristics) and even Java hasn't got that right either.

route on host (rfc5549) is the best strategy for host mobility if your fabric is designed around that idea and you embrace ECMP, stateless networking and bipartite topologies. You need to track state to do that without changing the FIB.

yea we did this in the UK during the 2011 riots, not technically protesting but same as this law.

so I believe this initiative is problematic but I feel activists are shooting themselves in the foot talking about privacy and conspiracy. The issue is lock-in -- the authority is mandating compatibility with proprietary Palantir feature sets (for example approach to federation and access control) and use at a local level (outside NHS-E) as a cost saving measure (the cost saving part means replace not dual operate 2 platforms). This is signaled strongly here and in the FDP procurement.

these leaders are a reflection of the voting population in this and those respective countries.

turns out we don't have that many smart people in the UK

research has found a bigger correlation with indoor time in early years than computers