I still don't see the appeal of Auto Mode. If I don't trust the model to run arbitrary commands, how much additional trust can be gained by an additional non-deterministic system. The point is that there are some failure modes which are completely unacceptable. Auto Mode lowers that risk but doesn't remove it. An unacceptable failure is still possible.
Non yolo mode where you manually approve each command is literally pure security theatre. No human on earth has the patience to discerne the huge stream of commands agents run.
So even if false negatives are technically zero, having 100% false positives is not acceptable.
So a fundamental premise is you have to have filtration. Where only very few things are surfaced for humans to look
The first option which a decent number of harnesses do is being able to setup an auto approve list. Like `ls` is fine, `cat` is fine etc.
Now maintaining this list is in and of itself a giantic pita. But the real issue is that it still has way too many false positives. Fundamentally it comes down to the halting problem where you can't really include important things like `bash python <<PY` and what not which agents like to use. But regex can't solve the halting problem to figure out if the Python is safe.
So naturally the next best option is to use an LLM. Which isn't that stupid because even if its non deterministic at least it can dramatically reduce the false positives from the regex auto approve list.
I think saying "nobody has the patience" is assuming that everyone is using agents for hours on end to do enormous tasks. I tend to prompt claude to do smaller, more contained tasks that I understand the entirety of, in which case manual review is straightforward and actually preferred, because I can stop it before it makes a change I disagree with.
once you have built up a trust in a model/harness/setup/workflow and you’ve been hitting enter for months straight then Auto Mode is appealing.
i found initially it would hang on some commands (moreso on subagents and moreso when first rolled out) and then roll past the hung command sometimes forgetting about it entirely. which wasn’t cool.
but after dialing in subagent permissions and probably some updates from anthropic Auto Mode is great. i still don’t always do it. but after reading this article just switched my current 12 panes to Auto.
Are there any plans to sandbox the agents? I suspect it would be fairly trivial on macOS using sandbox-exec. This looks very useful but I hesitate to run yolo mode agents (even with clear instructions) on my machine.
I have been impressed with Ente products and customer service. It's good to see they are growing. That said, revenue information is not entirely helpful in a vacuum. I'd be more keen on seeing profit (even at a lower timeline resolution). What's the average cost for taking on a new customer? What's the retention/turnover? Etc. That said the products are great and I'd recommend them to anyone.
I wish I could say the same. I tried Ente the other day, to see how it compares to Immich, but it was very hit-and-miss. Face recognition for people never worked, no matter what I tried, for example.
I'm in the same boat. I had a so many comments under Immich videos about Ente that it made me wonder if they were just bots.
The client / server relationship with Ente is peculiar, and on my test dataset of about 1000 images did not perform at all. Face recognition, semantic search, etc, it was not in the same league as Immich tbh. (Also hi Stavros!)
Hey Alex! Yes, exactly, I get that they're a fairly different category with different tradeoffs (trustless server), but also I own the server so I don't need it to be trustless. If it worked well, the tradeoff would be smaller, but as it stands it's not worth it for me over Immich.
Comparing self-hosted immich and a truly E2EE hosted platform like Ente is really odd. There are massively different trade-offs and target customers between these.
Ente is a private hosted photo storage and sharing platform. Only trusted devices that can decrypt /encrypt would even be able to perform the kind of work to tag photos for face recognition, and in this category I'm not aware of any hosted alyermative that actually does this, and they do a _pretty good job_ considering the limitations of E2EE.
Immich, Google Photos, Apple Photos- these platforms don't have to work around the fact all user data is opaque ciphertext. I imagine you could extend the capabilities of machine learning to a trusted node (similar to what you can do with Ente Desktop, but with more capable models), but they still have to work around any of that metadata being visible to them as well.
This takes somewhat trivial problems and makes them entirely non-trivial to solve. Self hosted platforms that don't provide an E2EE story have significantly less headache to deal with, especially since most people justify it by "it's running on my server so it's okay". And I get that, you can generally work around this issue with disk-level encryption and you're all good. It's just not the same product at all.
I agree it's not. It's just that, for someone who doesn't need the trustless architecture, it doesn't make sense to go with Ente right now, because it doesn't work as well. I could overlook the missing features if the ones that are there worked, but face recognition just got stuck for me and never finished.
Here is my use-case for a trustless architecture: there are parts of my families digital life which I would prefer to continue working as much as possible, even if I'm temporarily or permanently unavailable. "Pay this amount of money monthly/yearly for this to keep working" is a straight forward concept for 99.999% of the population. "Poke around in docker debugging logs if this stuff stuff" is not.
And, if stuff has to be hosted by third parties for me, I prefer it to be encrypted.
Who knows what the problem could be - if you want to invest time into this, you could try collecting some debug logs to see if it indicates anything. For me and my family it works pretty well (~300G of data across 10s of thousands of photos / videos). Face recognition / AI search works reasonably well. It's not perfect, but I'm happy to keep our photos under our control rather with some big tech company which could decide on a whim to erase all of my data (or even worse, report me to the police and then erase all of my data).
Of course I also dump all of the data about once per week on a NAS, just in case :)
I really enjoyed and agree with the majority of the article, but this was my nit as well. My hatred of vacation planning is often the reason I don't go on more vacations. It seems like automating a task that is experienced by the individual as completely monotonous ( and only affects that individual) would be a great example of something worth handing off to a text generator.
For me there’s a lot of risk in vacationing in a new area I have no idea about. ChatGPT helps me here.
It all comes down to people who have comfort in their own workflows and it takes mental load to change it. And then find reasons to work backwards to justify not liking AI.
I read though the GitHub readme but I'm still unsure what "new" this brings this brings to the table. It seems like a thin wrapper over existing tools. Since Microsoft rarely deprecates and removes anything, this feels like just another unnecessary complexity layer.
I'm not a traditional app dev on Windows though, so I'm likely missing something. For those of you who are more familiar, what about this are you excited about?
It is, but it's like "dotnet new" templates: a means of getting to a working minimal setup that jumps through the Microsoft hoops for you. MSIX and Package Identity are definitely headaches to get set up.
Yes, Package Identity was the bane of my existence for a while. It's wild how MS gated important APIs behind it and then made it so difficult+impractical to work with (I honestly think this is a large part of why WinRT was mostly ignored by developers).
I've moved on, but this looks like it will at least fix some of the paper cuts.
It's a CLI for especially web developers that prefer CLI tools over graphical installers and needing to have big IDEs installed like Visual Studio.
It's another small nail in the coffin of the monolithic Visual Studio, which is maybe exciting if you are rooting against Microsoft's paid products from their DevDiv and for their free-to-start options like VSCode.
