So this is like the case awhile back where the Waymo engineer joined Uber, brought a bunch of documents and got caught doin' a corporate espionage...but bigger. This is someone allegedly (with...
So this is like the case awhile back where the Waymo engineer joined Uber, brought a bunch of documents and got caught doin' a corporate espionage...but bigger. This is someone allegedly (with evidence) orchestrating a whole ring to do it.
OpenAI also instructs new hires on how to avoid scrutiny when they leave Apple. For example, Mr. Tan warns them not to tell Apple that they have taken jobs at OpenAI, so they can stay at Apple as long as they can. After his own departure, Mr. Tan improperly retained or obtained an internal Apple managers' document marked "Need to Know" that describes security procedures for employee departures. Messages left on Apple-issued work devices show that Mr. Tan and his OpenAI colleagues have been sharing this document with new hires before they give notice to Apple of their departures, previewing Apple's security protocols. Unsurprisingly, Apple's investigation has found a pattern by employees who depart for OpenAI of taking steps to evade the security processes intended to protect Apple's confidential information.
There are more quotes from the filing in the article, which have more allegations. And Apple followed that up with legal preservation letters sent to 40 former employees, who are presumably implicated.
This probably explains why Apple is shifting to Gemini models for things...and discovery is going to be nuts. I bet it'll make software companies start asking some questions about using OpenAI tools on their code bases.
Does anyone have a sense of how OpenAI are doing? I only have anecdotal everyday stuff to back this up, but it seems like most people I know who are frequent AI users have switched to Claude. Most...
Does anyone have a sense of how OpenAI are doing?
I only have anecdotal everyday stuff to back this up, but it seems like most people I know who are frequent AI users have switched to Claude. Most everyone else is using Google's integrated AI chatbot.
Is the sense that OpenAI are really struggling backed up by any data, or is that just the particular bubble that I'm in?
From a security and corporate espionage point of view OpenAI's program was laughably simplistic and known well outside the tech sphere, but given that I'm not American I thought they'd bought protection through bribes and cut-price government contracts. I guess maybe federally but not California-ly
They’re not doing great. Their financials leaked a couple of weeks ago, and several things stood out: the rate at which they’re burning money, the lagging revenue, and the accounting trickery they...
They’re not doing great. Their financials leaked a couple of weeks ago, and several things stood out: the rate at which they’re burning money, the lagging revenue, and the accounting trickery they use. The big one: lumping the subsidies and deals they give to consumers into the advertising spending category, since it’s a better looking cost for investors. They’re lumping so much spend in there, that they’re supposedly spending more on advertising than Coca Cola.
Another thing that happened just the past days & is a sign of them not doing great, is Oracle’s credit rating being downgraded to a step above Junk. I am not sure if you know how critical Oracle & OpenAIs relationship is, to both of these companies’ success.
Edit: and I forgot to add that after their financials leaked, it’s rumoured that OpenAI pushed their IPO into the next year. Which is IMO, not a sign of confidence in their business & ability to raise money, once their financials can officially get scrutinised by everyone.
I don't have insight into their financials, it's hard to see exactly what's happening with all the smoke from the burning piles of cash. But over the last year or so, for business use cases, and...
I don't have insight into their financials, it's hard to see exactly what's happening with all the smoke from the burning piles of cash.
But over the last year or so, for business use cases, and particularly coding, they've seemingly given up on leading and have instead been copying what Anthropic does in terms of harnesses and tooling. Their models themselves are usually a little behind Anthropic. In some situations a lot behind in my experience.
It's not an enviable position to be in when it's a tight race despite outspending the competition. Their head start is long gone.
I think they're well into profit on inference though. Which doesn't make up for their insane costs but it does make the business model look viableish. I think they've got some time yet before they need to worry about investors turning on them dramatically. Everyone is incentivized to hold on for the IPO.
And then at the same time, Chinese open weight models are getting dangerously close to the frontier. I'm surprised it hasn't caused a bigger correction. It's potentially really bad for both Open AI and Anthropic's profitability future, and therefore the public companies they're propping up.
I think I'd say they're in a precarious position rather than struggling. 5.6 Sol is a success from what I've seen so far, with a little help from the federal Fable shutdown. They're still undeniably in the top 2 with 3rd place quite a bit behind. As long as that doesn't change, and as long as the markets don't get spooked, they can probably keep the bonfire going for a while.
"profitable on inference" is only true if you ignore the fact that they offload lots of free trial and discount costs as marketing (not GAAP compliant, iirc). Unless you believe they spend...
"profitable on inference" is only true if you ignore the fact that they offload lots of free trial and discount costs as marketing (not GAAP compliant, iirc). Unless you believe they spend massively more on advertising than Coca-Cola!
That only matters if they booked the free trial as revenue and then deducted as marketing expense though. Otherwise, deducting as marketing expense is normal. Plus it doesn't impact the...
That only matters if they booked the free trial as revenue and then deducted as marketing expense though. Otherwise, deducting as marketing expense is normal. Plus it doesn't impact the calculation around profitability of inference. The only thing that matters to calculating inference cost is: is the price of computing a token less than the amount that a token is charged for. You can't hide that calculation by futzing around with customer acquisition cost numbers.
The GAAP-compliance part is irrelevant because most tech companies report both GAAP and non-GAAP as GAAP is not exactly the most accurate way to understand a business. They will report GAAP at IPO, with additional non-GAAP numbers. It's not weird because any analyst will be able understand the difference.
In many ways this isn't wrong, but I do feel like it's a bit reductive. 5.5 was a pretty meh model for sure, but 5.6 sol is legit really good. There have been some developer teams that moved off...
But over the last year or so, for business use cases, and particularly coding, they've seemingly given up on leading and have instead been copying what Anthropic does in terms of harnesses and tooling.
In many ways this isn't wrong, but I do feel like it's a bit reductive. 5.5 was a pretty meh model for sure, but 5.6 sol is legit really good. There have been some developer teams that moved off fable entirely once they got access to sol. And the codex computer use skill is way better than whatever Claude code can do. When I was having fable work on play date stuff, the first thing I had it do was write its own skill to control the play date simulator with AppleScript. That would not have been necessary with a model in codex.
Not to mention that OpenAI is the only lab that is focusing on token efficiency for their models. Take a look at the cost per task benchmarks. 5.6 sol is cheaper to run than sonnet 5, let alone opus or fable. Anthropic seems to build their models to burn tokens like no tomorrow. OpenAI is actually training for efficiency.
My colleagues and I have noticed that Claude Code seems to have gotten more expensive lately. Whether the cost per token has increased, or the application itself just likes to use more tokens,...
My colleagues and I have noticed that Claude Code seems to have gotten more expensive lately. Whether the cost per token has increased, or the application itself just likes to use more tokens, we’re not sure. But we’re actually able to use a significant portion of our individual budgets now. Previously, it felt difficult to really spend much.
I don’t know what’s going on internally, but I’ve been using my $20/month ChatGPT subscription quite a bit and I’m quite happy using GPT-5.6 for coding. Usually Terra. Since I don’t use Claude...
I don’t know what’s going on internally, but I’ve been using my $20/month ChatGPT subscription quite a bit and I’m quite happy using GPT-5.6 for coding. Usually Terra. Since I don’t use Claude Code, I can only use Claude at API rates and it’s too expensive.
Same here, 5.6 Luna has been more than adequate for what I need it to do. For now it's also been kind on my $10 Copilot Pro subscription + the $10 extra budget I gave it.
Same here, 5.6 Luna has been more than adequate for what I need it to do. For now it's also been kind on my $10 Copilot Pro subscription + the $10 extra budget I gave it.
So this is like the case awhile back where the Waymo engineer joined Uber, brought a bunch of documents and got caught doin' a corporate espionage...but bigger. This is someone allegedly (with evidence) orchestrating a whole ring to do it.
There are more quotes from the filing in the article, which have more allegations. And Apple followed that up with legal preservation letters sent to 40 former employees, who are presumably implicated.
This probably explains why Apple is shifting to Gemini models for things...and discovery is going to be nuts. I bet it'll make software companies start asking some questions about using OpenAI tools on their code bases.
Does anyone have a sense of how OpenAI are doing?
I only have anecdotal everyday stuff to back this up, but it seems like most people I know who are frequent AI users have switched to Claude. Most everyone else is using Google's integrated AI chatbot.
Is the sense that OpenAI are really struggling backed up by any data, or is that just the particular bubble that I'm in?
From a security and corporate espionage point of view OpenAI's program was laughably simplistic and known well outside the tech sphere, but given that I'm not American I thought they'd bought protection through bribes and cut-price government contracts. I guess maybe federally but not California-ly
They’re not doing great. Their financials leaked a couple of weeks ago, and several things stood out: the rate at which they’re burning money, the lagging revenue, and the accounting trickery they use. The big one: lumping the subsidies and deals they give to consumers into the advertising spending category, since it’s a better looking cost for investors. They’re lumping so much spend in there, that they’re supposedly spending more on advertising than Coca Cola.
Another thing that happened just the past days & is a sign of them not doing great, is Oracle’s credit rating being downgraded to a step above Junk. I am not sure if you know how critical Oracle & OpenAIs relationship is, to both of these companies’ success.
Edit: and I forgot to add that after their financials leaked, it’s rumoured that OpenAI pushed their IPO into the next year. Which is IMO, not a sign of confidence in their business & ability to raise money, once their financials can officially get scrutinised by everyone.
I don't have insight into their financials, it's hard to see exactly what's happening with all the smoke from the burning piles of cash.
But over the last year or so, for business use cases, and particularly coding, they've seemingly given up on leading and have instead been copying what Anthropic does in terms of harnesses and tooling. Their models themselves are usually a little behind Anthropic. In some situations a lot behind in my experience.
It's not an enviable position to be in when it's a tight race despite outspending the competition. Their head start is long gone.
I think they're well into profit on inference though. Which doesn't make up for their insane costs but it does make the business model look viableish. I think they've got some time yet before they need to worry about investors turning on them dramatically. Everyone is incentivized to hold on for the IPO.
And then at the same time, Chinese open weight models are getting dangerously close to the frontier. I'm surprised it hasn't caused a bigger correction. It's potentially really bad for both Open AI and Anthropic's profitability future, and therefore the public companies they're propping up.
I think I'd say they're in a precarious position rather than struggling. 5.6 Sol is a success from what I've seen so far, with a little help from the federal Fable shutdown. They're still undeniably in the top 2 with 3rd place quite a bit behind. As long as that doesn't change, and as long as the markets don't get spooked, they can probably keep the bonfire going for a while.
"profitable on inference" is only true if you ignore the fact that they offload lots of free trial and discount costs as marketing (not GAAP compliant, iirc). Unless you believe they spend massively more on advertising than Coca-Cola!
That only matters if they booked the free trial as revenue and then deducted as marketing expense though. Otherwise, deducting as marketing expense is normal. Plus it doesn't impact the calculation around profitability of inference. The only thing that matters to calculating inference cost is: is the price of computing a token less than the amount that a token is charged for. You can't hide that calculation by futzing around with customer acquisition cost numbers.
The GAAP-compliance part is irrelevant because most tech companies report both GAAP and non-GAAP as GAAP is not exactly the most accurate way to understand a business. They will report GAAP at IPO, with additional non-GAAP numbers. It's not weird because any analyst will be able understand the difference.
In many ways this isn't wrong, but I do feel like it's a bit reductive. 5.5 was a pretty meh model for sure, but 5.6 sol is legit really good. There have been some developer teams that moved off fable entirely once they got access to sol. And the codex computer use skill is way better than whatever Claude code can do. When I was having fable work on play date stuff, the first thing I had it do was write its own skill to control the play date simulator with AppleScript. That would not have been necessary with a model in codex.
Not to mention that OpenAI is the only lab that is focusing on token efficiency for their models. Take a look at the cost per task benchmarks. 5.6 sol is cheaper to run than sonnet 5, let alone opus or fable. Anthropic seems to build their models to burn tokens like no tomorrow. OpenAI is actually training for efficiency.
My colleagues and I have noticed that Claude Code seems to have gotten more expensive lately. Whether the cost per token has increased, or the application itself just likes to use more tokens, we’re not sure. But we’re actually able to use a significant portion of our individual budgets now. Previously, it felt difficult to really spend much.
I don’t know what’s going on internally, but I’ve been using my $20/month ChatGPT subscription quite a bit and I’m quite happy using GPT-5.6 for coding. Usually Terra. Since I don’t use Claude Code, I can only use Claude at API rates and it’s too expensive.
Same here, 5.6 Luna has been more than adequate for what I need it to do. For now it's also been kind on my $10 Copilot Pro subscription + the $10 extra budget I gave it.