Hacker News new | past | comments | ask | show | jobs | submit login
Anthropic hires OpenAI co-founder Durk Kingma (techcrunch.com)
162 points by coloneltcb 9 days ago | hide | past | favorite | 82 comments





Kingma is most notable for writing one of the most cited papers in AI. Actually one of the most cited scientific papers ever published, right up there with the transformers paper if not higher. "Adam: A Method for Stochastic Optimization" https://arxiv.org/abs/1412.6980

I remain astonished that Adam continues to be the most widely used optimizer in AI 10 years later. So many contenders have failed to replace it.


The number of papers that use it exceed the number of papers that cite it probably by 100x!

In my opinion, he is more notable for inventing a variational autoencoder.

I mean, there’s only so many ways to optimize a black box

(delete)

That's an architecture, not an optimizer. You can probably use ADAM with KANs. I think you latched on to the transformers in the sentence, but Kingma did not invent those.

> Anthropic’s CEO, Dario Amodei, was once the VP of research at OpenAI, and reportedly split with the firm after a disagreement over OpenAI’s roadmap — namely its growing commercial focus.

So he's now the CEO of Anthropic, a company selling AI services?

Claude is amazing, and we use it's Teams plan here at the office extensively (having switched from ChatGPT since Claude is vastly better at technical material and adcopy writing).

But, Anthropic definitely has a commercial motive... no?

I'm not saying a commercial motive is a bad thing - hardly... but this quote seems to be odd given the circumstances.


Pivoting from "for all mankind" to "all for myself" would make me deeply uncomfortable, too. The change from one position to the other, not either position in any absolute sense, is the concerning part.

This is also a great point. I ranted at length about this when the OpenAI news broke last week, but to cut it short: it's a little troubling to see the company founded on the ethos "for-profit AI work is incredibly dangerous" transition to a for-profit AI firm openly engaged in an arms race. Not just engaged, inciting...

https://web.archive.org/web/20230714043611/https://openai.co...


> Pivoting from "for all mankind" to "all for myself"

Isn't the former already a red flag?


No it’s good to try to build tech that helps people. Doesn’t mean such declarations need to be taken at face value, but being baseline-cynical is generally unwarranted, undesirable, and uninteresting.

The baseline stance for tech startups is wanting to solve a problem in the world and profiting from the value that provides. And thankfully, most of the time, those motives don't conflict. Even a mundane business like my local grocery store solves the problem of a curating a selection of food producers, buying in bulk to ensure a sustained sustained supply at a reasonable price and make it available to me close to my home. That is a tremendous value! And for that they make their markup. They aren't necessarily solving other social problems like food scarcity or maximizing nutrition or whatever, and instead focus on what their customers want to buy, those that can pay for it. But there still is a meeting in the middle of value being created.

I know a certain Nigerian prince who would wholeheartedly agree with such broadly encompassing statements.

Cynicism doesn't necessarily protect you from getting scammed, but it does absolutely prevent you from accessing any upside there is to be had in the world :)

The upside being people flocking from MLM to DeFi to LLMs like headless chickens while I watch in amusement?

The only downside for me is having been involved in all these projects and knowing enough to innovate. At least I do try to warn people before we proceed.


I think that the approach that Anthropic is taking to governance is a little different than "all for myself", it's worth having a read of https://www.anthropic.com/news/the-long-term-benefit-trust

It sounds like they are least trying to build on the notion of being a public benefit corporation, and create a business that won't devolve into chart must go up and to the right each quarter.

Time will tell of course, OpenAI was putatively started with good, non-profit intentions.


Does anyone actually believe any of this horsesh*t?

You're absolutely correct that they're a for-profit firm, but you're missing that they were founded specifically over safety concerns. Basically it's not just "commercial motive" in general, it's the sense that OpenAI was only paying lipservice to safety work as a marketing move.

For example, here's their research mission: https://www.anthropic.com/research

And an example of one of their early research focuses, Constitutional AI: https://arxiv.org/abs/2212.08073


I don't think it's odd.

* Acting in accordance with declared motivations is a demonstration of integrity.

* Acting towards hidden motivations that oppose your declared motivations is deceptive action.

Honest people don't want to lead and be responsible for deceptive action, even if the action is desirable.

For these types of people, it is often better to leave a place that requires them to active deceptively in favor of one that will let them operate with integrity.

Even if the end goal is the same, eg: to make money.


At least Anthropic is honest about their intentions though. That would be enough for me to leave OpenAI. Hey you want to commercialize it sure but don’t hide behind lies.

> I'm not saying a commercial motive is a bad thing - hardly... but this quote seems to be odd given the circumstances.

It only seems odd to you because you are reading much more into it than he's ever said, like "AI should never be commercialized under any circumstances and it is impossible to do so correctly". Then yes, it would be hypocritical. But he didn't say that and never has; and Anthropic thinks they are doing it right.


Also Anthropic, even though it has the same conflict of interest as OpenAI, seems to be addressing safety concerns more earnestly. They are listening to their internal team and innovating in the space (especially in a practical sense, not just future AGI concerns). Massively funding (or at least announcing too) and then ignoring a team is much worse.

I think the whole situation where they got some serious investment from SBF and then he got indicted pushed them into commercialising their tech so they could have more standard sources of funding

As I understand the main disagreement between OpenAI and Anthropic is exactly how much and what is censored.

I'm not sure as to what Anthropic means when they mean safety. I remember them doing good, non-censorship work in this field, but I also pay for ChatGPT instead of Claude because Claude is just so censored and boring.

Where is Claude more censored than ChatGPT?

For my use-cases, mostly erotica.

Can you go into more depth why you believe Claude is better @ creating adcopy

In my experience it is better at not sounding like an LLM wrote it, even without being directed to not sound like an LLM. It's better able to find and maintain the desired tone (playfulness, silly, professional, a mixture of, etc) with minor prompting. It also seems better at understanding your business/company and helping craft adcopy that's on-message/theme.

We used ChatGPT's Teams plan too with GPT4, but were sold on Claude almost immediately. Admittedly we have not used GPT4o recently, so we can't compare.

With technical information, Claude is vastly better at providing accurate information, even about lesser-known languages/stacks. For example, it's ability to discuss and review code written in Gleam, Svelte, TypeSpec and others is impressive. It is also, in our experience, vastly better at "guided/exploratory learning" - where you probe questions as you go down a rabbit hole.

Is it always accurate? Of course not, but we've found it to be on average better at those tasks than ChatGPT.


As somebody new to Claude, can anybody give me tips for how to optimally use Claude as opposed to habits formed with ChatGPT? For example, my main concern is the limiting of messages over a given time period, even for paid accounts. I have often used ChatGPT for very specific questions/answers, but sending a large collection of "drill-down" follow up questions can burn through my Claude messages pretty quickly? Is it as simple as composing longer, more fleshed out prompts to begin with (addressing follow ups ahead of time?) or is this where something like Projects helps? Thanks for any feedback!

Projects can make it worse because I think the Claude rate limiting is token based. If I fill up a project to >100k tokens, I get rate limited much faster.

Highly recommend using the API with a good client. It's cheaper and limits are way higher. I rarely hit my limit which is just a billing limit I impose intentionally.

My tip would be to use openai instead.

There's an arms race. Openai was ahead. Then anthropic was ahead. Now gpt4o and o1 are better again. This may change in a few months.

I'll miss the projects feature though.


Gpt4o and o1 are absolutely not better at coding than Claude. They're noticeably worse at following complex, detailed prompts (e.g. a detailed prompt for it to create an application), and often forget details during the conversation (e.g. will revert to an old version of some code it already refactored). o1 is better than Claude for Leetcode hard style coding problems, but the majority of coding work isn't about that, it's about correctly implementing a spec. Plus even o1 will still often fill code with "implemention here" comments, in spite of being explicitly asked to provide a full working implementation.

I write a lot of code with both in multiple languages. o1 is astronomically better than Claude. It's not even a competition.

I have a totally different experience.

I'm porting a medium sized project (40k loc) from iOS to Flutter and I couldn't be more happy with my setup with Claude. Every time I hit the Pro plan limit and I have to resort to ChatGPT the work that I have to put in to manually fix the code easily triples.


It's exciting to wonder where Anthropic will be in like 5 years time with this incredible momentum

Let's catch up about this together while we're waiting in the breadline.

Where society will be...

Always trippy for us apocalyptic optimists to read coverage about safety concerns and consolidating power in AI firms that reads exactly like these companies have been reporting for 20 years for smart phone apps, B2B SASS battles, and hospitality industry schemes. Reminds me of today's articles on the escalating war involving at least one nuclear power mentioning the Dow Jones as the fourth bullet point, but on an even larger and more ridiculous scale.

Godspeed to Anthropic! Hopefully they can be a force for good, despite the various deals with the devil that they've taken. They've lost so many safety and e/acc people that I was getting dubious, but they certainly are staying in the fight.

Shame they're already for-profit... But don't worry, they Pinky Promise to be For The Public Benefit :)


> Shame they're already for-profit... But don't worry, they Pinky Promise to be For The Public Benefit :)

Anthropic is legally a Delaware Public Benefit corporation so it's written into their corporate governance.

How effective that governance will be at balancing the public benefit with profit remains to be seen, but it's a lot more than just a pinky promise.


Thanks for the correction! I was mixing up "Public Benefit Corporation", which is a legal offering by state governments, and "B-Corp", which is a non-profit that certifies wholesome for-profit firms like the Tillamook Dairy Cooperative.

I'd stand by the general assertion that it's little more than a pinky promise because they merely have to "balance" the concerns according to "any reasonable person"--an extremely weak-seeming obligation to this non-lawyer--but it's certainly much more impactful than I thought, namely:

  Sections 365 (b) and (c) provide broad protection to directors of public benefit corporations against claims based on interests other than those of stockholders
https://www.legis.delaware.gov/BillDetail?LegislationId=2235...

Good on you, Anthropic! In this specific case I believe in the director(s) a lot more than I believe in the shareholders ethics-wise, so it seems like a perfect choice. They can always fire him/them I suppose, but truly catastrophic AI risks would move faster than that, anyway.


Yes, this is the real legal accomplishment of B-Corps!

By law and precedent a C-Corp’s only obligation is to shareholders, thanks to a case from almost a century ago: https://en.m.wikipedia.org/wiki/Dodge_v._Ford_Motor_Co.#:~:t....

A B-Corp was the first and somewhat successful attempt to create a legal framework where company executives are allowed to work on behalf of all their stakeholders without it creating an automatic basis for a suit.

Generally, people who care very deeply about a thing bring a higher ethical standard than any regulatory body can impose.


Can anybody suggest a good open source (GUI or terminal-based) app for chatting with Claude Sonnet for those who have API keys? I use those for a neovim plugin to chat given the context of codebase, but I would also like an ability to have a regular chat like in the web interface?

https://pypi.org/project/llm/

I use Open WebUI for when I want a website with some more features than a terminal provides.


LibreChat


Mira Murati next?

I hate the penchant Claude has for abstracting everything into one line functions

Left door, right door

As much as I dislike OpenAI’s ongoing shenanigans and disdain for their own customers, I tried to sign up for Claude last week.

Turns out that Anthropic’s signup flow has been silently broken for months for Firefox users: https://old.reddit.com/r/ClaudeAI/comments/1bq06yz/phone_ver.... You get the SMS verification code, and you can enter it, but you get a barely visible “Invalid verification code” error message followed near-instantly by a refresh of the page. I reached out to support, but like many others, heard nothing back.

This barely-disguised contempt for what should likely be their most valuable power-user base suggests to me that a lot of the recent departures from OpenAI are being driven by push instead of pull, and I’m not convinced that Anthropic will remain a competent competitor in the LLM arms race long-term.


> This barely-disguised contempt for what should likely be their most valuable power-user base suggests

Yikes. I am a long time Mozilla supporter, active user of Firefox since before it was Firefox, and former Mozilla employee, but this comment is pretty crazy.

Firefox is well below 3% market share, and is essentially a niche browser at this point - it sucks when I run into sites and services that aren't supported by Firefox, but I don't assume that it's contempt for me as a Firefox user. I simply assume that I, as a power user, have opted to use an alternate tool that has features that are compelling to me, and I certainly don't expect every business out there to prioritize my use of a niche tool.

I learned a long time ago that while power users can be an effective avenue for building a market for niche products, they also end up being some of the most problematic users, because of the assumptions that power user needs should be placed above the regular users. It's fine to want to be catered to, but it's not really great to assume malice when you aren't - it shows contempt for the prioritization of the limited resources they have available.


Firefox is going to be massively disproportionately more popular among LLM users, the issue has been happening to people for over half a year, the error message is terribly misleading, and from a CS POV, it should be relatively easy to flag and route customers around the known issue.

I’m not salty just because they don’t support the browser; you’re totally right that that’d be an unreasonable take, but it’s not the one I’m trying to make.


> Firefox is going to be massively disproportionately more popular among LLM users

Source for this claim?


For what it's worth I signed up for Claude on Firefox without issues several days ago. I'm not saying the issue isn't real, but it isn't universal on the browser.

Yeah - I run the ESR version of Firefox, which is probably the root of it, but wide prevalence of people hitting the same issue for so long + the total radio silence from support is what really threw me. At least OpenAI pretends to care by responding with AI-generated pseudohelp…

> total radio silence from support

the new customer reality, courtesy Google etc..


I'd like to add to this.

I signed up and paid for credits to access their API last weekend.

All requests still get rejected saying I don't have sufficient credit. This is despite their dashboard saying that I do indeed have the requisite credits.

No response despite reaching out to support.

Don't think I have been treated this indifferently by any other service in recent times.


Lol - talk about making a mountain out of an anthill

You’re not wrong, but I’m more leaning on the anecdote to make a broader case that the big walled-garden players (Google/OAI/Anthropic) all kinda suck in similar ways.

I.e. - I think Anthropic is seeing a boon right now not because they’re doing things right, but because the competition is doing them worse.


Anthropic is a joke of a company. They have all these heavyweight hires, but they make logins so difficult for the user that no chat user in their sane mind would want to use Anthropic. It's as if Anthropic doesn't actually want people using their service.

They routinely keep logging me out, also always making me wait for an email confirmation code just to login every time, and it's sickening.

They also promise API credits but then don't actually give any.


I'm a chat user, hopefully with a sane mind, and their login process might be annoying but doesn't stop me from using it. I get logged out once every 2-4 weeks maybe?

My experience has been opposite. I'm using the API. I hit usage tier limits so sent them a request to upgrade my tier. Heard back within an hour and upgraded to the highest tier. I had another request also which was answered same day.

That's good but my complaint is not about the API at all. It is more foundational in that most people would want to use the API (for applications) only if the web chat works well for them, which it doesn't.

"More foundational" == their login method lol

The chat on the website is not their main product... They're selling access to their models to enterprises. As sickened as you are by having to log in to the chat, that's not an indication of their success at training and marketing high-quality models (the only real competition to OpenAI at this point).

> The chat on the website is not their main product

Guess what: enterprises are made of people. People like to try things out. If people are not happy with something for their personal use, they most definitely are never going to recommend it to their employer. This is why OpenAI wins. It is in fact one of the factors that sets apart a hyper-successful product from a wannabe.


Literally every single person I know who’s building anything in AI (dozens of early stage founders) prefers Claude over GPT at this point.

The AI emperor will not be the one who has the most consumers logging into product.com to use the chatbot.

Compound this with OpenAI’s continuous shedding of, as far as I can tell, every credible researcher… I find your position quite hard to believe, even accounting for the hysterical tone.


> Literally every single person I know who’s building anything in AI (dozens of early stage founders) prefers Claude over GPT

I don't believe this at all. I am not here to argue that Claude is a worse model, only that Anthropic is a worse company.

> The AI emperor will not be the one who has the most consumers logging into product.com

Your point only goes to show how much Anthropic hates its end users.

> OpenAI’s continuous shedding of, as far as I can tell, every credible researcher

OpenAI has zero trouble hiring great talent. As I see it, they lost a lot of dead weight that had no interest in bringing AI to the masses, but had an agenda of their own instead.


“Your point only goes…”

Huh? How so? Sorry not even clear what your complaint is… is it the Firefox (3% market share) login bug? The Claude chat experience has been superior for a while now, and Projects and Artifacts make it 100x so.

Good at hiring and bad at retaining is much worse than the reverse, especially for long-lived R&D projects.


> what your complaint is

It helps to read. It's noted in the original comment. It has nothing whatsoever to do with Firefox, as it manfiests only on the Anthropic website.


Ah that's right, the "sickening" pattern of emailing a login code. Not a fan of it myself, but boy is this dramatic.

It's that, but it's more. They promise API credits to try out the API, but then refuse to give any credits. Overall, it's as if they don't care even a little bit for individual customers.

Shredding you say?

Seems to me that the only difference between Anthropic and OpenAI is that Anthropic was for-profit from day one and OpenAI is from day yesterday. I pay for both, and pretty sure they will do everything they can to take as much money from me that they can get away with.

This shouldnt be news.


The for profit from the start is true, but also, Sam's shenanigans really irks me as a customer, the slights of hand etc. I get the sense he would mislead the public on threats and risks of AI to benefit OpenAI and the government, to centralize and monopolize powerful models.

What if your city police force or volunteer firemen switched to for-profit?

I think that is the crux of the matter.


The US is unique in how many public services it has. Other countries have private firefighter services; that just means the city has a contract with them. It doesn't mean they burn your house down and charge you for it.

maybe a better analogy would be:

People set up and fund a public bus system that has coverage for all neighborhoods, rich or poor, distant or close.

And then after the bus system is up and running, the bus system manager decides transportation is important! He IPOs the bus system, and changes all the routes to money-making routes with cost optimized (higher) fares.


See, one reason it doesn't matter if the public directly owns a bus system is that you can just pass laws telling them what routes they have to run.

But also, a money-making bus system would mean people actually use it. Nobody uses most American bus systems because they don't go anywhere, and they're slower when they do because they aren't popular enough to replace car traffic.


It depends on what would actually change.

If they started paying volunteer it would tell me the town has more money not they will do a bad job now that they are paid.



Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: