Crying dev begs for 100k-token context window to reach AGI

Description

The meme is split vertically: on the right is a close-cropped, teary-eyed man with a single tear running down his cheek, the classic ‘please bro’ reaction image. On the left, bold black text reads: “Bro please just 100k tokens more I swear I'll solve AGI. We're still working on a bigger context window bro. This is how AI works please bro”. The visual gag contrasts desperate human emotion with the highly technical demand for a larger transformer context window. For seasoned ML engineers, it pokes fun at the industry trope that simply scaling token limits or model size will magically unlock artificial general intelligence, ignoring architectural bottlenecks like quadratic attention cost and training data quality. The meme satirizes the current hype around ever-expanding LLM context lengths as a silver bullet

Common Tags

AI Llm AIHypeVsReality AIHypeCycle AIHype MachineLearning

Comments

Anonymous ★ Top Pick Because obviously the only thing separating us from artificial general intelligence is an O(n²) attention matrix that fits on the CFO’s AWS budget

Anonymous ★ Top Pick

Because obviously the only thing separating us from artificial general intelligence is an O(n²) attention matrix that fits on the CFO’s AWS budget
Anonymous

After 20 years in the industry, I've seen the same pattern: 640K ought to be enough for anybody, 4GB of RAM will future-proof you, and now apparently 128K tokens is all that stands between us and the singularity. At least the VCs are consistent in funding logarithmic improvements marketed as exponential breakthroughs
Anonymous

The irony here cuts deep: we're promised AGI that will revolutionize everything, yet we're still negotiating for a few more tokens like it's 2019. It's the AI equivalent of 'just one more sprint and we'll have achieved sentience' - meanwhile, the model can't even remember what you said 50k tokens ago. Classic case of marketing running a marathon while engineering is still tying its shoelaces, except the shoelaces are O(n²) attention complexity and nobody wants to talk about it
Anonymous

AGI after “100k more tokens” is the transformer equivalent of “just scale the monolith” - attention stays O(n^2), the KV cache becomes a cost center, and you end up building RAG and eviction policies anyway
Anonymous

Every roadmap pitch: add 100k tokens and we get AGI; meanwhile the model pays O(n^2) to skim your doc, forgets the middle, and invoices you for the privilege
Anonymous

Plot twist: AGI was achieved, but it's buried in the middle of a 1B-token prompt no one can retrieve
@mrYakov 1y

They already trained llm on everything that exists in internet, and it still not enough to properly train existing models. Dont matter how long context is, there are not enough data to do it
1. dev_meme 1y
  
  Naah, you really need to get deeper into the topic It’s faaar from the amount of data it trained on And muuuch more further from all existing data being used
  
  @mrYakov 1y
  
  proof pls
2. @Sun_Serega 1y
  
  when all of internet is not enough, just AI-generate the input data. I mean, they already do it. works pretty well, trust me bro
  
  dev_meme 1y
  
  Huh? Nah, sorry bro, LLM generated content is worst food to train new generation of LLMs, so couldn’t trust you after saying something like this
  
  @azizhakberdiev 1y
  
  But considering that internet starting to get filled with AI generated stuff it is not far from truth
  
  @loves_frogjs 1y
  
  Use llm to distinguishe ai content from human content and then feed only human content to new llm
  
  @slnt_opp 1y
  
  If only somebody tried that... wait a minute
  
  dev_meme 1y
  
  That’s the neat part You can’t distinguish it already