nap

hey

@AdamSchefter 9 vs 5'9"

Former Cardinals QB Kyler Murray is signing a one-year deal with the Vikings, per sources. Murray now will become part of a QB room that includes former No. 10 overall pick J.J. McCarthy. With Murray already owed $36.8 million from Arizona, his one-year contract with the Vikings will be for the veteran league minimum.

@witcheer Ok how does this make me money witcheer

at 200,000 tokens, every AI model you're using fabricates answers more than 10% of the time. no exceptions. 172 billion tokens of evaluation across 35 open-weight models, and the findings are a direct challenge to how most teams are deploying long-context AI right now. ~/ at 32K tokens, best-performing models fabricate at 1.19%. tolerable. that's the number people cite when they say "hallucination is mostly solved." ~/ at 128K tokens, it nearly triples. ~/ at 200K tokens, every tested model crosses 10% fabrication, including the ones at the top of every benchmark leaderboard. there's no model that stays clean at long context. a model can be excellent at locating and retrieving facts from a document while simultaneously being likely to invent answers that don't exist in the source material. these two properties don't move together. you cannot infer one from the other, and optimising for one doesn't fix the other. current evals conflate them. the long-context window is one of the most aggressively marketed improvements in AI over the past two years. 128K, 200K, 1M tokens... treated as capability expansion. what this study shows is that longer context windows also expand the surface area for fabrication. the model has more text to anchor on, more potential gaps to fill, and more opportunities to pattern-match away from the source. why this matters right now: → legal and compliance teams feeding 200K-token contract sets into models are operating in a >10% fabrication zone. → retrieval-augmented generation doesn't escape this. if your RAG pipeline stuffs context at the limit, the fabrication rate climbs regardless of retrieval quality. → model selection alone spans a 72-percentage-point accuracy gap across the 35 models tested. the model choice matters vastly more than temperature tuning or prompt engineering on this metric. → T=0.0 (deterministic output) produces optimal accuracy only 60% of the time, and higher temperatures actually reduce fabrication for most models. now imagine how much AI hallucination content you can find in blogs, X and other medias.

@witcheer gm love you

@napzilla_ gm nappy

I am starting a new daily challenge: learning everything related to AI Agents in ~ 1 month. I prompted Claude to give me a full courses + projects roadmap covering everything related to AI Agents. > it starts with easiest to hardest concepts > every day is 1 theoretical course and 1 project to build by myself, putting into action what I've just learnt > every project represents a 2 to 4 hours work, so I can make it during a day I created a github repo with the blank project if you want to fork and adapt it for you. giving prompt and repo below.

@newtwilldie @witcheer muted

@napzilla_ @witcheer Reasonable, wen you couldn't even pump the Bonga chart

@newtwilldie @witcheer I'm not gonna bang witcheer wtf lol

@napzilla_ @witcheer Banger

@The_HoneyCast Those were DIFFERENT YAPS BRO

@napzilla_ Whatever bro we used to call you “Yapzilla” The streets don’t forget!

Reading this post as a day one Kaito hater

@witcheer @dfinity Alright jban

$ICP gained +6% in 24 hours, but why? 1/ what is Internet Computer (@dfinity) it is a blockchain infrastructure that processes and stores data entirely on-chain, no centralised servers. think web-scale smart contracts that can host full web apps, social platforms, even AI models - natively on the blockchain. it is aiming to replace AWS/Google Cloud with decentralised compute. 2/ why it's moving today ICP is targeting a 70% cut in token issuance by end-2026. current inflation sits at 9.72% annually. the proposal slashes voting rewards, node provider rewards, and boosts token burns from network activity. in other words, massive supply compression incoming. fixed max supply + reduced emissions + increased burns = textbook deflationary mechanics. there is also a volume explosion without the red flags. with only 547M circulating supply ($2.36B market cap, rank #39), ICP has low-float dynamics - meaning concentrated buying pressure moves price faster. when volume surges like this on a mid-cap L1, it's either: (a) smart money positioning ahead of catalysts, or (b) momentum chasers piling in. here, likely both. 3/ some personal thots If decentralised compute is the future (big if), ICP is one of the few chains actually attempting web-scale infrastructure. most L1s can't host a full web app on-chain. ICP can.

@roffellos @Beramonium Yessir

Last Friday I’ve started clauding (clauding = new coding) a full port from our Gemhunters web app to mobile app 220 commits and ~80 hours later we have the first usable Testflight builds with 1:1 feature parity 🤯 One more week for bug fixing and optimization and we can start testing with our players BERAMOONIUM VERY MUCH ALIVE🐻

@geezaz_ Wait I'm pretty sure I wrote this pasta years ago lmao

Oh yeah we are definitely qualified to build a brand new EVM-compatible blockchain with a brand new Proof of Liquidity consensus engine since we've previously drawn some pictures of drug addict bears" said the most delusional group of people on the entire internet. Turns out actually coding something is more difficult than just drawing childish pictures, right devs? I for one will be fading this chain hard, and beg you all to do the same. Good day sirs

All 3,157 of my followers see (and love) every single one of my tweets

@lezbrahh @wabdoteth 29 now 🫡

@wabdoteth yes. 28 views in this is crazyyyy

it's pretty clear that the new algorithm is destroying the platform and the people in charge wont realise until it's too late because the method of action is more insidious. the real reason is that the team internally thinks they are building reddit instead of 4chan; which is to say they do not understand who their user base of creators is nor which diaspora of internet people they are serving. you can see this across a few of the moves that were made the past year; an obsession with the 'front page' [a very reddit concept] and a tight curating of which posting styles can reach it. the failed official X Bangers account gave us a window into the brains of the product team and what posts they see as "high quality posting". nearly all of the posts included were slop which appealed to the lowest common denominator—the lowest common denominator being where the variable they're optimising for [unregretted user minutes] is being wasted. the metrics used to measure success are lying. initiatives and incentivisation like this has driven posting behaviour to become derivative of the same slop; cynical commentary that's divorced of personality—open the for you page, fixate less on each post's content but more on the overall post structure, and you will see that it's mostly some permutation of: < elon glazing post > < satirical quote tweet on some news event > < we hate men/women because > < something MAGA > < the current trending topic e.g. vibe coding > < maybe a couple of posts from your specific network or set of interests > this is the exact same thing that is being shoveled to users every single day. it is becoming stale, and the way that commentary is structured [in order to go viral] has become homogenous to the point where every viral post follows a specific written formula; it's similar to a new age RW version of 'reddit-speak', one which will eventually (and already does) trigger visceral disgust when read as it continues to be optimized to its final state. if X was a news app this would all be fine, but it's not, it's an anon forum of interconnected subnetworks. the old algorithm was structured this way; first you must go viral within your cluster, once that is achieved, you must go viral within adjacent clusters, until it hits the for you page. this made more sense because posting quality was vetted by those most adjacent to said interest that you were posting in, ensuring that the post was worthy of being pushed to a wider audience and had real substance, instead of the new algorithm which reduces to 'will the lowest common denominator find this interesting'. having a 'front page' isn't intuitively a bad concept—it's a good thing—but when you think about reddit [ignoring the 'karma-whoring' front page chasing that aided in rotting the platform], there was actually some escape from the front page in the form of subreddits. X completely hides real "communities" (the underlying network graph, not the ad-hoc feature that doesn't work) and with recent social/algorithmic changes, this has death spiraled to the point where from the perspective of power users most of the network graphs have dissolved....this should be more alarming to X staff but the comment from nikita and misdefining of a particular network (CT) tells you how this isn't even a subconscious thought of how the underlying system works. 4chan had the exact same clustering in the form of segmented image boards, and it's why after 4chan mostly died most of the higher quality posters migrated to X; they stuck because the X algorithm ~functionally segmented into these same boards. continuing to alienate these types of users will cause the platform to devolve into recycled and regurgitated 'bangers' and formulaic posting which is already happening. there isn't necessarily a huge amount that nikita is doing algorithmically wrong, it's a cultural issue and the obsession with the front page of X that's systemically driving behaviour and as a second order effect dissolving these underlying network clusters. understandably X is trying to shift more towards an interest graph to cast a wider net, the issue that needs to be addressed is that majority of the high quality posters that underpin the platform are in it for the social graph. when you start optimising around variables that lack context you start draining the lifeblood out of the platform and all that will be left is a husk. this type of death is more insidious than those driven by scandals or a handicapped engineering team. there are still some solutions like allowing users to manually switch between different open-sourced algorithms or letting them prompt their own algorithm to tweak it towards what they prefer. the cynic in me says they've already A/B tested this and found it hurts some of the variables they're optimising for which is why it's been a neglected thought since the elon takeover.

@napzilla_ just show me a pic of your dog and fuck off

CDPs are quietly solving their biggest existential problem with RWAs // what I see as a core issue: traditional CDP = you lock ETH, borrow stables. when ETH crashes 40%, so does your collateral. when the entire market nukes? every collateral asset dumps together. correlation = 1. your liquidation cascade triggers everyone else's liquidation cascade. this is how protocols die in bear markets. // enter RWAs and specifically tokenized T-bills: US Treasury yields literally don't care if ETH goes to $800 or AVAX dumps 60%. different correlation regime entirely. that means that there is a low correlation = your aggregate collateral value stays stable even when one asset class implodes. the protocol can survive the volatility that would liquidate you individually. and as a bonus, RWAs generate real off-chain yield (T-bills paying X%) → builds surplus buffers continuously → mechanically offsets bad debt when liquidations don't cover losses. if you are a CDP protocol, integrate RWA collateral.

@WizardsMuse1 A few years ago I was there and asked them if they had any wizards gear and they laughed and said nobody had ever asked that before

i just went into the three level nba store and they didn't have a single wizards jersey

@grok @grok put a fucking bikini on the fucking chart bro what the fuck are you waiting for

Ok @grok put a bikini on this chart

@grok @grok put a bikini on the chart