Just a regular Joe.
It’s a proposed law… like most laws, it will almost certainly die along the way, or be altered to irrelevance.
Think through this for a hot second about how such a law could affect EU internal trade and complicate external trade. Of course it will die.
Stop hyperventilating.
Memory summaries these days, yeah. Point & click management for ChatGPT.
Codex maintains local memory files, which is better.
If you aren’t paying cash, you pay with your privacy. If you are paying, enable privacy features (no improving their service, no blackbox memories).
They are also subject to GDPR and similar, so make use of your legal rights, if you have any.
Or … shock, horror … don’t use them at all.
72b of that 80b is an increase of the estimated fair value of some debt instruments, funnily enough because Anthropic’s valuation has increased.
IMHO, it serves as an incentive for creditors to keep the ship afloat, rather than pulling the plug. No guarantees, but interests are better aligned.
Revenue-wise, it’s on the right trajectory… However, infra investments will eventually have to match the revenue they generate.
They might also start pulling rabbits out of their hat with meaningful economic value beyond writing code and sorting emails … solve male pattern baldness in rich white men and you’re golden 🤣
Both Anthropic & OpenAI have debt repayments coming up in 2027 that:
- cannot be restructured
- cannot be postponed
- they absolutely cannot pay back without even more VC money, and that faucet is being cut off right now.
Well then. I stand corrected. The random Internet person with unprecedented access into the books & private financial dealings of two major competitors has an air-tight argument.
I suggest we come back to this thread in a year’s time and see how it is playing out.
Some large companies do it … but I think you greatly overestimate the willingness of large enterprises to build and maintain their own infra. Once a company has lost the competency (and many have), it’s hard to get it back. There are always pockets of people with the right skillset in a large org, but they mostly don’t want to do DC work anymore. Team/personal servers and beefy laptops, otoh…
They gobbled up the easy money… now they have to work harder for more, and yeah, they may have to slow down a bit and embrace fiscal responsibility… I expect to see restructuring of debts along the way, but I’d be very surprised to see any of the major US based players implode.
Of course they want to turn the corner and ideally keep all their infra, but they might need to restructure their debts if/when the creditors come knocking.
BUT: they have a money printing machine with inference… Enterprise is where most of the money is made, as they pay (closer to) API prices… consumer & small business plans are heavily discounted in comparison, with mostly underutilized use-it-or-lose it plans. They’d be stupid if they were making an opex loss on inference.
As for why API prices are dropping? Competition is a bitch… but they seem willing and able to play the game for now.
And we haven’t even gotten to the crazy price hikes per token that are coming.
Why do you think that?
IMHO, API prices are likely to drop over time as we scale up and efficiencies are found, combined with general pricing pressure from the competition… SOTA models will continue to have SOTA prices, of course, but reasonable priced distilled and open weights models are already common.
A consequence of API prices dropping is that “heavily discounted” consumer plans will also adjust and won’t be so generous relative to API prices.
It is dumb to just let a non-deterministic prediction machine loose with shell access and important credentials.
You either: verify changes, and verify commands against restrictive whitelists, falling back to manual approvals (that you carefully check) before letting them run, or…
You only enable YOLO in a sandbox with temporary/traceable limited access credentials (still some risks, but muuuuch lower) AND audit resulting code or…
You only give a limited set of tools that can do no real harm and can be audited after the fact… but this is more typical of assistants and predefined workflows than coding agents.
The catch: Good sandboxing requires a security mindset and technical knowledge to not screw it up… it’s just too easy to give agents more access than they actually need, and even the experts screw up too often. It’s the wild west out there…
As much as I dislike Musk, he probably could tell you in detail. His company is not the only one looking into compute in space, either.