Researchers watching Apple abandon reasoning LLMs before even pushing first commit

Description

Meme image in POV format: three young men in semi-formal attire stand shoulder-to-shoulder, their faces blurred for anonymity, staring directly at the camera as if witnessing something baffling. Behind them is a collage of densely packed research-paper pages - charts, tables, and academic figures clearly visible - evoking an arXiv wall of recent breakthroughs. Overlaid text in bold white Impact font reads, at the top, “POV: watching Apple giving up on” and at the bottom, “reasoning LLM without even trying.” The humor plays on Apple’s reputation for polished but closed-ecosystem products juxtaposed against the open research community’s push for chain-of-thought and reasoning-capable large language models. Senior engineers will catch the irony: while the background papers suggest rapid progress in transformers and CoT prompting, Apple is depicted as exiting the race before producing a single experiment, highlighting the tension between corporate risk aversion and bleeding-edge ML research

Common Tags

Apple Llm AIIndustryTrends AI AIHypeVsReality MachineLearning IndustryIrony

Comments

Anonymous ★ Top Pick Turns out multi-hop reasoning doesn’t fit in the Neural Engine’s budget - Cupertino is shipping ‘return nil’ as a feature

Anonymous ★ Top Pick

Turns out multi-hop reasoning doesn’t fit in the Neural Engine’s budget - Cupertino is shipping ‘return nil’ as a feature
Anonymous

Apple's approach to reasoning LLMs is like their approach to USB ports - they'll wait until everyone else has figured it out, then claim they invented a revolutionary new way to not do it
Anonymous

Watching Apple abandon reasoning LLMs is like seeing a senior architect reject microservices after only trying a monolith with REST endpoints - technically, you haven't even started exploring the design space. The irony is that reasoning capabilities in LLMs are heavily dependent on prompt engineering, few-shot examples, and chain-of-thought techniques, yet here we are witnessing a Fortune 500 company apparently giving up before implementing basic prompting strategies. It's the enterprise equivalent of declaring 'Kubernetes is too complex' after only running `docker run` once
Anonymous

Apple saw the test‑time compute bill for ‘reasoning’ and chose battery life - so we get RAG + summarization on a 4‑bit decoder; Think Different, just not at inference time
Anonymous

Apple's LLM verdict: memorization masquerading as math - every MLOps engineer's 'emergent ability' nightmare confirmed
Anonymous

Apple’s alignment strategy for reasoning seems simple: if test_time_compute > battery_budget, return NotAUseCase();
Sure Not 1y

>2018-2019: yooo bruh what if there is tech that can convert human language into code >>openai and chatgpt appears >fuck, go back.
1. Sure Not 1y
  
  Lets see if this time it works.
@deerspangle 1y

Apple actually giving a shit about releasing a product with quality and purpose? Wild times
1. @Box_of_the_Fox 1y
  
  Nah, they simply ignored llms and are fucked now. Siri was quite amazing at a time but for some reason they decided not to improve it
  
  @deerspangle 1y
  
  Have you checked out their paper?
  
  @Box_of_the_Fox 1y
  
  Not really, which one?
  
  @deerspangle 1y
  
  It's linked further up in this thread, about the limits of "reasoning" LLMs
  
  @Box_of_the_Fox 1y
  
  Looking at abstract and how short actual paper is, I'd assume it was quickly put together to explain to shareholders why apple is so far behind in the subject. It doesn't say anything new. In case of apple it is bad because of Siri. It needs LLM of some sort and if apple can't create one then it's on them. They have more than enough cash and resources to create something that's half decent
  
  @deadgnom32 1y
  
  but why would THEY need a reasoning LLM? don't see any practical use case.
  
  @Box_of_the_Fox 1y
  
  To understand what their users are saying to the device and do the reasoning of what to do with it. That's what things like Siri and Bixby do and what Gemini is getting pretty good at. But no worries apple fanboys, you'll get that in 5 years and will believe it's a feature that no one ever had. Unless OpenAI gets there first as they seem to be working on that xd
  
  @Box_of_the_Fox 1y
  
  It's quite remarkable how quickly apple looses control over their ecosystem and no one seems to notice
  
  @deadgnom32 1y
  
  you don't need a reasoning LLM for this.
  
  @Box_of_the_Fox 1y
  
  You don't know what LLMs are then :p
  
  @deadgnom32 1y
  
  I am a AI Researcher, certified
  
  @Box_of_the_Fox 1y
  
  And I am a fox on the internet
  
  @TheFloofyFloof 1y
  
  But are you a certified fox on the Internet?
  
  @Box_of_the_Fox 1y
  
  How do I get such certification D:
  
  @deadgnom32 1y
  
  it's just, people nowadays tend to use overcomplicated tech for simple tasks, which have way better and stable solutions with less complex tech. that's what scientists are now talking about. but Sam altman promises, 1 more B parameters, 1 more $B. one more GPU cluster, and AGI will be here. promise fingers crossed
  
  @Box_of_the_Fox 1y
  
  Simple attention based models aren't complicated. You can do amazing stuff in prolog but it's not enough here
  
  @deadgnom32 1y
  
  yes but why would you need a reasoning, where nobody is observing this reasoning process? there are models capable of producing an output more directly and efficiently.
  
  @deadgnom32 1y
  
  more over, reasoning models are statistically more error prone. it's oftentimes funny to observe, how it thinks X, and misses this X in the output a second after
  
  @deadgnom32 1y
  
  whereas a more simple model, without reasoning and with 2-3 times less parameters produces a correct result faster and more accurate. 💁‍♂️
  
  @Box_of_the_Fox 1y
  
  Wait, so apple made an article that they don't need reasoning and now apparently everyone are amazed about it? What's next? Article that they don't need to compete with Stax in audio equipment?
  
  @deadgnom32 1y
  
  I don't know about everybody and apple. I had this opinion long ago. it's oftentimes like this — a long struggle with LLM can be replaced with a simple mathematical model with a couple of dozens parameters, not billions, and get far more accurate results with less energy consumption. you can run it basically on a fridge
  
  @nolemocius 1y
  
  could you please give a couple specific examples? I would need this to look smarter in pointless internet arguments with other dogs like me
  
  @deadgnom32 1y
  
  I am. will need some time to find a paper from the last conference with good examples.
  
  @Box_of_the_Fox 1y
  
  There was some research showing that SVM is still the most popular model and compared to stuff like LLM you could describe it as simple mathematical model. There are also examples showing that you could get away with prolog in simple chat bots. Historically there was similar development with computer vision about which you can read in mobilenetv2 white paper
  
  @deadgnom32 1y
  
  look for: Legal Chunking: Evaluating Methods for Effective Legal Text Retrieval 275
  
  dev_meme 1y
  
  Literally single most important reason why latest Gemini are so good - their reasoning/CoT are next level
  
  @deadgnom32 1y
  
  I am no apple user, never liked their soft, and tech seams overpriced to me. but to achieve same results. you can use a more basic LLM without reasoning capability.
@Bjastkuliar 1y

Tbh they do make some solid points in the paper. Is it new discovery? Definitely no. Is it surprising that a tech giant states it? Totally.
1. @Johnny_bit 1y
  
  Lnk to paper for those out of the loop plz?
@einbetungzahl 1y

https://news.ycombinator.com/item?id=44203562
@Johnny_bit 1y

Of multiple thankings

Use J and K for navigation

Researchers watching Apple abandon reasoning LLMs before even pushing first commit

Description

Categories

Common Tags

Comments

More like this

Do Nothing With AI, Win: Jacked Tim Cook

When One Gigaflop Became A Weapon

Giving Codex a Break: IQOS Vape Connects to the Mac