mlcruz

I feel like your writeup could be a little stronger if you made it clear how important is hitting those recall targets.

I also do believe search quality is recall dominated, but the link between the metrics and results are not as clear as they could have been. I think you have a stronger sell here than it looks at first glance.

Congratulations to the team tho! Those are very nice results

“Please don't use Hacker News for political or ideological battle. It tramples curiosity.”

Can we keep at least this place clean of this kind of discussion?

Both yours and the other guys side are having this same discussion on every single platform possible. Do you guys really need to do it here?

I think we keep getting used to intelligence improvements and the goal posts are always moving forward.

I remember that close to an year ago i felt the same way about opus 4.5. And even if you can easily run ds or qwen locally to get very similar results to that, we moved to 5.5 being the new target.

Deepseek can do 90% of the job but a claude reviwer/spec writer is a brutal improvement.

What worked well for me was a custom version of Open Web Ui with some customization to spawn an exe.dev instance for each new chat. I can just work on my phone, deploy stuff for development purposes on an easy to share way etc.

It is not that rare, the epidemic is just focused on the south region right now.

Porto Alegre metropolitan area is having a huge outbreak. My girlfriend is a vet and has been dealing with new cases multiple times a week.

Many of our friends also got it (it is very hard to not get scratched when handling with cats in pain).

It is a really really shitty, painful and hard to treat disease, requiring multiple months of treatment. It is very painful but usually not letal for humans and cats that are in the earlier stages and get treatment.

However it is absolutely lethal for populations of wild and stray cats, as it is very infectious and 100% lethal unassisted.

If its just a single session, without too many parallel agents, fable on xhigh lasts an entire session without hiting linits.

Sadly since fable usually works comfortably for 10-20min at time without human input, i end up juggling at least 3 other agents and it lasts me about 2 hours.

If i have a really hard problem or big refactor, i use workflows. This consumes the entire session quota in about 45 minutes.

I really enjoyed the read. Not everything needs to be some sort of utilitarian information density optimized reading piece.

Keep up the good work!

I have been using deepseek via deepinfra, afaik they provide no data retention. Im probably going to deploy the full model on their infra instead of paying credits at some point, so far the experience has been pretty good

My workflow is quite similar. I try to write my prompts and supporting documentation in a way that it feels like the LLM is just writing what is in my mind.

When im in implementation sessions i try to not let the llm do any decision making at all, just faster writing. This is way better than manually typing and my crippling RSI has been slowly getting better with the use of voice tools and so on.

> I think you misunderstand what I'm saying. I'm not really referring to data systems at all, I'm referring to context on what problems are actually being solved by a business. LLMs very clearly do not model outcomes that don't have well-defined textual representations.

Yeah i misunderstood your point, i completely agree with what you are saying.

I honestly do not believe that strategy, decision making and other real life context dependent are going to be replaceable soon (and if it does, its something other than llms).

> I'm not sure that I agree with white collar jobs being done for, not every process has as little consequence to getting it wrong as (most) software does.

Maybe im too biased due to working in a particularly inefficient domain, but you would be surprised how much work can be automated in your average back office.

Much of the operational work is following set process and anything out of that is going to up the governance chain for approval from some decision maker.

LLM based solutions actually makes less errors than humans and adhere to the process better in many scenarios, requiring just an ok/deny from some human supervisor.

By delegating just the decision process to the operator, you need way less actual humans doing the job. Since operations workload is usually a function of other areas, efficiency gains result in layoffs.

> we have a means of giving them full business/human levels of context

Trust me, this is a work in progress. Right now most corporations do not have their data organized and structured well enough for this to be possible, but there is a lot of heat and money in this space.

Imo, What most of the people that are not directly working in this space get wrong is assuming swes are going to be hit the hardest: There are some efficiency gains to be won here, but a full replace is not viable outside of AGI scenarios. I would actually bet on a demand increase (even if the job might change fundamentally). Custom domain made software is cheaper as it has ever been and there is a gigantic untapped market here.

Low complexity to medium complexity white colar jobs are done for in the next decade through. This is what is happening right now in finance: if models stopped improving now, the technology at this point is already good enough to lower operational costs to the point where some part of the workforce is redundant.

Hi Phil,

Your article is great! As someone who's working in this space, your points just improved our presentation and selling a lot. We have been talking with C level finance executives about building semantic layers, and i can confidently say that the way you presented the value proposition of the context layer is going to improve our conversion rates.

Thank you so much! This is one of the best analysis i have ever heard about the subject.

Really? This sounds very interesting, as the USA is the place that i would expect to have the most malls.