Is there a startup that takes models like this, and effectively gives you a secure setup, where you have (a) a mobile app that (b) talks to some giant machine that only you have access too.
If a 10K computer could run this, it may be worth it to have a "fully on prem" version of ChatGPT running for you.
Another reason they could have built this was by listening to their users. I do believe lots of people are spinning up agents in their workplaces, and managing yet another set of api keys is probably annoying for Tailscale's customers. This feels like a great solution to me.
Pressure to service larger customers to capture higher revenues is inevitable for Tailscale given the scale of VC funding, valuation, and operating costs involved.
Trying to be all things to all people will inevitably dilute focus, and it’s understandable that OP might be looking at this sub-product and wondering where the value is for their use cases.
They’re probably not the only ones questioning whether they’re still part of Tailscale’s core ICP (ideal customer profile), either.
yes this inevitably happens to companies that can't grow infinitely, you pivot to enterprise because you can sell to one person that has the equivalent spend of thousands... it really is unfortunate for the individuals
I have a secret manager, why would I want tails ale involved in the management of secrets, they are a networking company
Tails ale is not a company I see being involved in my core AI ops. I don't need their visibility tools, I already have LGTM.
Tailscale should focus on their core competency, not chase the gilded Ai hype cycle. I have sufficient complaints about their core product that this effort is a red flag for me. To do this now, instead of years ago, shows how behind the times they are
They're not a networking company, they're an access control company. Their original product is based around networking, and now this new one is based around AI access and metrics.
This product isn't about managing and distributing API keys, it's about managing and distributing access to these services throughout the org. In fact, it's more about being able to avoid managing and distributing API keys, which is IMHO even better.
The first I heard of them was they were the company around WireGuard, a networking technology.
We recently brought them into the stack to manage said access, it has been painful, aiui their configuration is not intuitive (not the one working on it). I suspect any further expansion will be a big ask after the dismal experience. I certainly don't trust them to manage my secrets and access afterwards. I haven't even found an enjoyable DX talking point in either my personal or professional usage either
> They're not a networking company, they're an access control company.
This is like Ripping saying they are not an HR company, they are an access control company. I got into this very argument with them on a sales call looking for a payroll provider. They wanted to manage the keys to everything, I don't trust them to safely guard access to my cloud projects, nor is it something I even want my HR/payroll company even considering doing. This new product sounds like TailScale was the keys to the kingdom and I sure as hell am not giving it to them after the disappointing rollout of their established networking technology
Not to mention that storing the API keys on a developer machine (or distributing them to a developer machine) is the first step towards a developer's API keys getting leaked or exfiltrated. With this approach, the developer never has the API key on their machine at all (and you don't have to rotate or invalidate the key when they leave).
There's a set of common needs across these gateways, and everyone is building their own proxies and reinventing the wheel, which just feels unnecessary.
~All of our customers at Oso (the launch partner in the article) have been asking us how to get a handle on this stuff...bc their CEO/board/whatever is asking them. So to us it was a no-brainer. (We're also Tailscale customers.)
For those curious about the question: "how well does GPT 5.2 build Counter Strike?"
We tried the same prompts we asked previous models today, and found out [1].
The TL:DR: Claude is still better on the frontend, but 5.2 is comparable to Gemini 3 Pro on the backend. At the very least 5.2 did better on just about every prompt compared to 5.1 Codex Max.
The two surprises with the GPT models when it comes to coding:
1. They often use REPLs rather than read docs
2. In this instance 5.2 was more sheepish about running CLI commands. It would instead ask me to run the commands.
Since this isn't a codex fine-tuned model, I'm definitely excited to see what that looks like.
The only time I spent outside of the video was to deploy to Vercel. I made a bunch of speedups in the video, but didn't cut anything. The total time was about 2 hours.
I mentioned it in the post, but there was definitely some hand holding towards the end, where I don't think a non-programmer would have succeeded
I am sorry this made you feel depressed. I think there are some positives to consider too though:
1. More people that wanted to make games can.
Thanks to unreal engine, you don't need to be a Tim Sweeney level-expert to make compelling games. I see LLMs as another abstraction in the same spirit.
2. You get more leverage
The more abstractions you have, the more you can do with less. This means less bureaucracy, more of a chance to make _exactly_ what you wanted.
I understand how the craft changes underneath you, and that can feel depressing, but if we see it as tools, I think there's lots of good ahead.
I agree that this is not how shooting is implemented in a production competitive shooter, but that wasn't the goal of the post. We wanted to test how far models could go, and I think this current version is the limit in a 2 hour time window.
GPT’s differentiator is they focused on training for “thinking” while Gemini prioritized instant response. Medium thinking is not the limit of utility
Re: overparameterization specifically Medium and High are also identically parameterized
Medium will also dynamically use even higher thinking than High. High is fixed at a higher level rather than leaving it to be dynamic, though somewhat less than Medium’s upper limit
Has Google made anything as impactful as Gmail and Chrome since those acquisitions? Stagnant doesn't mean unprofitable.
Google videos was a thing, it died. Google had an awesome modular phone project (Ari?), it died too. I can imagine they could have done something like M-series apple chips and an actual modular phone and phone OS that was superior to Apple products.
If a 10K computer could run this, it may be worth it to have a "fully on prem" version of ChatGPT running for you.
reply