Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Reflections on OpenAI

calv.info

Post reply on HN
Print view View original

231–240 of 446 posts

Re: Reflections on OpenAI

#231

Post by stickfigure »

rrrrrrrrrrrryan wrote: post #121
hinterlands wrote: post #52

It is fairly rare to see an ex-employee put a positive spin on their work experience. I don't think this makes OpenAI special. It's just a good reminder that the overwhelming majority of "why I left" posts are basically trying to justify why a person wasn't a good fit for an organization by blaming it squarely on the organization. Look at it this way: the flip side of "incredibly bottoms-up" from this article is that…

> There's no Bond villain at the helm. It's good people rationalizing things. I worked for a few years at a company that made software for casinos, and this was absolutely not the case there. Casinos absolutely have fully shameless villains at the helm.

Interesting. A year ago I joined one of the larger online sportsbook/casinos. In terms of talent, employees are all over the map (both good and bad). But I have yet to meet a villain. Everyone here is doing the best they can.

Re: Reflections on OpenAI

#232

Post by lossolo »

ivape wrote: post #204
echoangle wrote: post #187

Earlier quoted context omitted.

> There was nothing hypothesized that next-token prediction and scale could show emergent properties. Nobody ever hypothesized it before it happened? Hard to believe.

Someone else can confirm, but from my understanding, no they did not know sentiment analysis, reasoning, few shot learning, chain of thought, etc would emerge at scale. Sentiment analysis was one of the first things they noticed a scaled up model could generalize. Remember, all they were trying to do was get better at next-token prediction, there was no concrete idea to achieve "instruction following", for example. W…

The Schaeffer et al. "Mirage" paper showed that many claimed emergent abilities disappear when you use different metrics, what looked like sudden capability jumps were often artifacts of using harsh/discontinuous measurements rather than smooth ones.

But I'd go further: even abilities that do appear "emergent" often aren't that mysterious when you consider the training data. Take instruction following - it seems magical that models can suddenly follow instructions they weren't explicitly trained for, but modern LLMs are trained on massive instruction-following datasets (RLHF, constitutional AI, etc.). The model is literally predicting what it was trained on. Same with chain-of-thought reasoning - these models have seen millions of examples of step-by-step reasoning in their training data.

The real question isn't whether these abilities are "emergent" but whether we're measuring the right things and being honest about what our training data contains. A lot of seemingly surprising capabilities become much less surprising when you audit what was actually in the training corpus.

Re: Reflections on OpenAI

#233

Post by jjani »

ivape wrote: post #207
jjani wrote: post #50

> The thing that I appreciate most is that the company is that it "walks the walk" in terms of distributing the benefits of AI. Cutting edge models aren't reserved for some enterprise-grade tier with an annual agreement. Anybody in the world can jump onto ChatGPT and get an answer, even if they aren't logged in. There's an API you can sign up and use–and most of the models (even if SOTA or proprietary) tend to quickl…

OpenAI hides their model's CoT (inference-time compute, thinking) Probably because Deepseek trained a student model off their frontier model.

And the same thing could very easily happen to Anthropic, yet they choose not to hide it.

Re: Reflections on OpenAI

#234

Post by ZYbCRq22HbJ2y7 »

alwa wrote: post #46
bhl wrote: post #3

> The Codex sprint was probably the hardest I've worked in nearly a decade. Most nights were up until 11 or midnight. Waking up to a newborn at 5:30 every morning. Heading to the office again at 7a. Working most weekends. There's so much compression / time-dilation in the industry: large projects are pushed out and released in weeks; careers are made in months. Worried about how sustainable this is for its people, gi…

If anyone tried to demand that I work that way, I’d say absolutely not. But when I sink my teeth into something interesting and important (to me) for a few weeks’ or months’ nonstop sprint, I’d say no to anyone trying to rein me in, too! Speaking only for myself, I can recognize those kinds of projects as they first start to make my mind twitch. I know ahead of time that I’ll have no gas left the tank by the end, and…

I think any reasonable manager would appreciate that sort of interest in a project and would support it, not demand it.

Re: Reflections on OpenAI

#235

Post by ZYbCRq22HbJ2y7 »

Rebelgecko wrote: post #59
bhl wrote: post #3

> The Codex sprint was probably the hardest I've worked in nearly a decade. Most nights were up until 11 or midnight. Waking up to a newborn at 5:30 every morning. Heading to the office again at 7a. Working most weekends. There's so much compression / time-dilation in the industry: large projects are pushed out and released in weeks; careers are made in months. Worried about how sustainable this is for its people, gi…

How did they have any time left to be a parent?

They were showered with assets for being a lucky individual in a capital driven society, time is interchangeable for wealth, as evidenced throughout history.

This guy is young. He can experience all that again, if it is that much of a failure, and he really wants to.

Sure, there are ethical issues here, but really, they can be offset by restitution, lets be honest.

Re: Reflections on OpenAI

#236

Post by shermantanktop »

stickfigure wrote: post #231
rrrrrrrrrrrryan wrote: post #121

Earlier quoted context omitted.

> There's no Bond villain at the helm. It's good people rationalizing things. I worked for a few years at a company that made software for casinos, and this was absolutely not the case there. Casinos absolutely have fully shameless villains at the helm.

Interesting. A year ago I joined one of the larger online sportsbook/casinos. In terms of talent, employees are all over the map (both good and bad). But I have yet to meet a villain. Everyone here is doing the best they can.

Every villain wants to be the best villain they can be!

More seriously, everyone is the hero of their own story, no matter how obvious their failings are from the outside.

I’ve been burned by empathetically adopting someone’s worldview and only realizing later how messed up and self-serving it was.

Re: Reflections on OpenAI

#237

Post by danenania »

maxnevermind wrote: post #215

What I really wanted to know if OpenAI(and other labs for that matter) actually use their own products and not just casually but make LLM a core of how they operate. For example: using LLM for coding in prod, training/fine-tuning internal models for aligning on the latest updates, finding answer etc. Do they put their money where their mouth is, do LLMs help with productivity? There is no mention of it in the article…

I don’t know, but I’d guess they are using them heavily, though in a piecemeal fashion.

As impressive as LLMs can be at one-shotting certain kinds of tasks, working in a sprawling production codebase like the one described with tight performance constraints, subtle interdependencies, cross-cutting architectural concerns, etc. still requires a human driving most of the time. LLMs help a lot for this kind of work, but the human is either carefully assimilating their output or carefully choosing spots where (with detailed prompts) they can generate usable code directly.

Again, just a guess, but this my impression of how experienced engineers (including myself) are using LLMs in big/nontrivial codebases, and I’ve seen no indication that engineering processes at the labs are much different from the wider industry.

Re: Reflections on OpenAI

#239

Post by danenania »

dcreater wrote: post #158

This is silicon valley culture on steroids: I really have to question if it is positive for any involved party. Codex almost has no mindshare and rightly so. It's a textbook also ran, except it came from the most dominant player and was outpaced by Claude code on the order of weeks. Why go through all that? Instead what would have been a much better scenario is openai carefully assessing different approaches to agent…

Codex is quite different from Claude Code. It’s more similar to Devin.

Maybe you’re thinking of the confusingly named Codex CLI?

Re: Reflections on OpenAI

#240

Post by whamlastxmas »

vonneumannstan wrote: post #78

>Safety is actually more of a thing than you might guess Considering all the people who led the different safety teams have left or been fired, Superalignment has been a total bust and the various accounts from other employees about the lack of support for safety work I find this statement incredibly out of touch and borderline intentionally misleading.

This is not true and it’s weird hyperbole misinformation
Post reply on HN