Magnus

@BourbonCap Everything AI coming from microsoft has been so incredibly mid, always late and worse than competitors, why would you want them in control?

$MSFT still owns 27% of Openai Unpopular opinion: i want Openai to fail as a business so Microsoft can eat the whole thing. If nothing happens, thats fine as well

@mtbomb @thdxr Cloud infrastructure is one of the most profitable, highest margin businesses (AWS, Azure…)

@thdxr This plan only works if the tokens you're selling are not a commodity. If several firms are offering tokens that are interchangeable, there won't be a ton of profit similar to web hosting

inference is very profitable and probably a good opportunity to understand some basic business math 1. companies buy long lived assets like GPUs. these are one time costs and the asset depreciates over time 2. once you own this asset, you can plug it in and produce tokens which you can sell. the cost of goods sold here can be very low and you might be making 90% margins at scale, this is why we say inference is profitable 3. then you also hire employees to do r&d work to improve your systems, come up with new models, expand the business if you add these 3 up you end up with $0. you're not producing a profit because the business is growing and you're reinvesting it all buying assets or r&d to meet demand if it's obvious to other people the business is working, you can raise money from them to accelerate all these numbers so they max out in 5 years instead of 25 so on paper you'll be "losing money" every year but that's because you want to make sure you lock down the opportunity before someone else the bigger your market is the bigger this burn can be because it's a function of potential so when you see these companies losing a lot of money it doesn't mean the whole concept of their business broken it's possible they misjudge and overinvest on 1+3 and will suffer some consequences but fundamentally 2 does work

@GrindstoneSEO Heard of em, but should be pretty easily detectable with a small scraper script? Could also check for topical relevance of the linking content using embeddings/NLP when scraping

@WasBruba Yeah, that's the link farm indicator metric. They're not really link farms but scraper farms designed to trap Googlebot on the network but Claude won't listen to me about that part.

Finished building this today, now running it through multiple iterations catching more and more edge cases that passed to the review file.

@GrindstoneSEO Makes sense yeah - I guess you are running with a dataset of sites spammers link to based on the wording? Probably the most fuzzy part of this but do you also look at linked sites from linking url I'd say thats an indicator

@WasBruba No single factor qualifies or disqualifies a link. It's a very complex scoring system that I developed through a LOT of trial and wrror.

@GrindstoneSEO Are you really running with no money keyword anchors? Havent been too engaged with link tactics recently but aint that filter legit links?

Who can spot the logic error?

@alxfazio You can just put running the linter into your agents.md dont know any llm that fails doing that after edits, what more could you want?

i expect native linting at llm write time soon. ruff is incredibly powerful

@badlogicgames This feels very notion todo listing - why spend time working on the tool when you could be using - you know - the software it produces lol

beautiful

@ninthalek @thdxr Yes you can do that ask it for „invariants“

@thdxr Did anyone try to introduce testing axioms for LLMs? Like: - Testing is the process of executing a program with the intention of finding errors. - A good test case is one that has a high probability of detecting an undiscovered error. - It is impossible to test your own program.

lmao this is maybe the craziest LLM written test i've ever seen

@realmcore_ Ah, yeah makes sense actually saw one in the wild I followed in my seo days x.com/hobo_web/statu…

@WasBruba No? They will build their own tools

All sales people will be technical

@realmcore_ Unironically @roon for this post x.com/tszzl/status/2…

Doing lit review for this blog post and I really do not want to do it. This is pretty awful. However, I will be including references to some of the most insightful people in the space at each level of the stack if you have anyone in mind, please let me know so I can evaluate

@BrendanFalk I think skills are mostly working pretty badly, personally landed on <thing>.<type>.md that .<type>. helps a lot to get agents to actually read that stuff ironically - then just properly guide in your main prompt/agents.md

Key takeaway from all the comments: Use nested skills. e.g. instead of separate skills for "create PDF" and "parse PDF", have one skill called "manage PDF" which then routes to the relevant sub-skills With good nesting, this can likely scale to 1000+ skills/sub-skills!

@lateinteraction Put dont use libraries in you agents.md or whatever - then let it iterate, I mostly tell codex to work in „slices“ and just chain „read agents.md“ prompts that gets it pretty well to implement stuff properly on a small scale

I still find it borderline stupid that coding agents seem inclined to use APIs or libraries in complex scripts before tinkering at small scale, as in bottom-up notebooks, to make sure they're modeling these APIs correctly. Who is responsible for this and what are they thinking.

@Gana_L_ @thsottiaux Switched to medium its much less prone to overthinking and didnt notice a difference for most task tbh, only using xhigh to plan complex tasks

@thsottiaux I'm also seeing insane "draining" when using 5.4 high. It uses like 3x more tokens than 5.3 xodex xhigh consistently I burnt through 50% of weekly usage in plus sub in under 3h While 5.3 xhigh couldn't even use 30% in half a day

We have found one issue that leads to some users seeing inconsistent usage across sessions but it is quite rare, affecting less than 1% of users. We are working on a mitigation and continuing the investigation. For the rest we are not seeing evidence of higher usage consumption other than the advertised token cost increase of GPT-5.4 being 30% higher than GPT 5.2 and GPT-5.3-Codex.

@realmcore_ (and not by the claude „production ready trust me bro“ criteria)

@realmcore_ This over decomposituon i mostly just define exit criteria for a task which obviously for something like this are basically endless - its more an experiment than anything else lol - but for things like crud apps and sane stuff you can get the model to call something done

I’m not sure RLM is the best way to do it right now… Some thoughts:

@realmcore_ Having mine build a distributed p2p storage system with gossip impl, where do you find limits? github.com/Mvgnu/maxio

you thought there was anything better than RLM? Currently writing a blog about this exact topic and why RLM could be peak but also where it breaks

Truly one of a kind @Apple the most basic stuff doesn’t work anymore, is anyone capable left at that company

Code Analysis Intensifies