AI assistant remembers exactly the wrong thing, proving ML still needs work

Description

The meme shows a minimalist chat screenshot on a white background. Grey user bubbles read “remember this location” and later “what did I ask you to remember”. White assistant bubbles, each prefixed with the four-color Google Assistant icon, reply: “Ok, I'll remember that”, then “I remember you told me.”, and finally “"This location"”. Below the chat, center-aligned black text proclaims: “Machine Learning is the future.” The humor comes from the assistant storing the literal phrase instead of the geocoordinates, illustrating natural-language understanding shortcomings and the gap between machine-learning hype and real-world behavior familiar to developers

Common Tags

MachineLearning AIAssistants AI AILimitations TechHumor AIHypeVsReality

Comments

Anonymous ★ Top Pick Our chatbot’s “context memory” is basically Redis.set(“remember this location”, “this location”) - but hey, the slide deck still calls it a transformer-based cognitive architecture

Anonymous ★ Top Pick

Our chatbot’s “context memory” is basically Redis.set(“remember this location”, “this location”) - but hey, the slide deck still calls it a transformer-based cognitive architecture
Anonymous

This is what happens when you deploy a stateless Lambda function and tell the PM it has 'memory capabilities' - technically it remembers everything you said, just not what any of it actually meant or referred to
Anonymous

This is a perfect example of why we still can't replace senior engineers with AI: it's mastered the art of technically correct answers that are completely useless. The assistant successfully stored 'this location' in its state management system while completely failing at reference resolution - essentially implementing a HashMap<String, String> when the architecture clearly needed a semantic knowledge graph with entity linking. It's the software equivalent of a junior dev who passes all unit tests but ships a feature that does exactly what the ticket said, not what the user needed
Anonymous

AI memory at scale: we built a billion-dollar NLP pipeline that reliably remembers the string "this location" - turns out the future of ML is a key-value store with really expensive autocomplete
Anonymous

When your assistant’s NLU treats deictics as strings, you get green unit tests and a memory store full of “this location” - eventually consistent with nowhere
Anonymous

When your transformer overfits to the prompt so hard it aces verbatim recall but fails zero-shot comprehension
@nuntikov 5y

Beta male: no, I don't want Google to know where I am at 😖😣😢 Sigma male:
1. @RiedleroD 5y
  
  Sigma males use 20 layers of proxies and GPS spoofers
  
  @nuntikov 5y
  
  Sigma male has nothing to hide, because there's nothing that can be used effectively against him, nothing he is ashamed of
  
  @RiedleroD 5y
  
  sigma male hides his information anyway, because he likes to annoy big tech
  
  @ZgGPuo8dZef58K6hxxGVj3Z2 5y
  
  They forgot that thare are at least 50 other google phones with wifi and bluetooth on that create a net of all devices. So google knows where you are even if you use gps spoofers
  
  Deleted Account 5y
  
  using pinephone
  
  @ZgGPuo8dZef58K6hxxGVj3Z2 5y
  
  I am not but ok
  
  Deleted Account 5y
  
  i mean that's the solution for google
  
  @ZgGPuo8dZef58K6hxxGVj3Z2 5y
  
  Oh well yeah that would really help
@mainfme 5y

Shitty Google
1. @yoyatayo 5y
  
  Its fake lol
  
  @mainfme 5y
  
  But Google still shitty
  
  @yoyatayo 5y
  
  U use what
  
  @feskow 5y
  
  duckduckgo.com
  
  @RiedleroD 5y
  
  I use ecosia.org - shitty search engine, but at least I feel like I'm doing good from time to time
@VlP_AI_TG 5y

It's big brain time

Use J and K for navigation

AI assistant remembers exactly the wrong thing, proving ML still needs work

Description

Categories

Common Tags

Comments

More like this

Stuffing Kimi K3 2.8T Into a 3090

One Prompt Ends an Eighty-Seven-Year Conjecture

The Alpha Model Skips Retrieval Entirely