I tried asking Fable 5 to identify the fungus in a picture I uploaded of one of my wife's plants. Apparently it thought I was trying to build a bioweapon. Opus answered it (yellow dog vomit fungus). Now I can spread the spores and take over the world!
I feel like the over safe aspect of the system will eventually back fire by doing stuff like "since humans always want to always destroy thing, they must be eliminated to stay on the guard rails". If thats how you align a system, its fundamentally wrong.
Just a personal anecdote but I have not hit any more thresholds or limits since switching to the MAX plan and so far, it's been worth it. But I do wonder how long even this will last...
I think subscription models are sustainable, but longer term, we should probably expect to see more prompt optimization happening in the providers inference pipeline. For example, unless you explicitly tell the agent or API to use a specific model, fronting the inference layer with a caching prompt classifier to determine which model to use, and automatically select the lowest cost model would probably already save alot of money (IDK if Claude/OpenAI do this on the backend, but several services I have worked on do some things like this to reduce costs of delivery customer facing inference at scale).
> fronting the inference layer with a caching prompt classifier to determine which model to use, and automatically select the lowest cost model would probably already save alot of money
Unfortunately, that doesn't work within a single session. The K-V cache of a model is intertwined with the model's configuration. Switching models invalidates the cache, meaning everything up to the point of the switchover is processed like a new, uncached input token.
Per Anthropic's pricing doc, an Opus 4.8 cache hit costs 50¢/MTok, while Haiku costs $1/MTok for uncached input.
Model selection works best if sessions are short and self-contained, particularly if the first few interactions can reliably classify the model need. That probably covers most 'support chatbot' use-cases, but it doesn't describe the kinds of heavy agentic automation that really chews through token budgets.
There is a definite financial incentive for people smarter than me to solve the problem, and I don't generally bet against businesses finding ways to reduce costs :)
> The K-V cache of a model is intertwined with the model's configuration.
I don't think this is true if you simply quantize the model or run it with fewer active experts? The underlying weights would stay the same. You could also play further tricks with skipping some of the model's middle layers outright, which works surprisingly well due to how skip connections are used.
I tried ultracode today on the max pro plan. An hour and a half in was all I lasted. Giant review on an entire six month old code base. It found 61 bugs, about ten were notable. Pretty impressed.
Ultracode destroys your limits and I have not found it to be worth it in the slightest, just fyi. I haven’t found any improvement over a local Claude code instance set to opus max.
You know the split is inevitable. Same as it ever was...
Whether that's Linux on your personal desktop and Windows on your work machine...
Oh and you built that desktop yourself, didn't you? But you can't even open the one at work or it's a violation.
GrapheneOS on your personal phone, and iOS on your work phone...
When this AI bubble crashes, we'll all be flooded with graphics cards no one else will want and all kinds of cool things will be built (are being built).
If you can stick it out a little longer you'll be fine. The tech you want to tinker with will be there.
Yes, in fact, during my days of red teaming, if I couldn't get info on an HVT (think celebrity), I'd go after their not-famous relatives[1] and friends, who generally had very bad opsec/persec. It was extremely effective.
1. An effective tactic is to friend relatives and friends on social media. From there, you either get to the HVT's data because it's set to viewable for "friend of friend" or you be patient, friend more of their friends and family and eventually friend the HVT directly, using your "connections" as social proof.
A very famous celebrity family was very susceptible to this tactic. After this project, they... tidied up their social media permissions.
It's a hardcoded default password, not a permanent backdoor. If I'm understanding the post correctly, the user changes it as part of the onboarding flow.
This is the way most apps work if they have a default password the user is supposed to change.
The device should ideally have some kind of secret material derived per device, like a passphrase generated from an MCU serial number or provisioned into EEPROM and printed on a label on the device.
Some form of "enter the code on the device" or "scan the QR code on the device" could then mutually authenticate the app using proof-of-presence rather than hardcoded passwords. This can still be done completely offline with no "cloud" or other access, or "lock in"; the app just uses the device secret to authenticate with the device locally. Then the user can set a raw RTSP password if desired.
This way unprovisioned devices are not nearly as vulnerable to network-level attacks. I agree that this is Not Awful but it's also Not Good. Right now, if you buy this camera and plug it into a network and _forget_ to set it up, it's a sitting duck for the time window between network connection and setup.
I agree that would be nice, but it also doesn't sound all that practical for a small vendor.
I used to sell a home networking device,[0] and I wouldn't do what you're describing. If there were an issue where the labels calculate the wrong password or the manufacturer screws up which device gets which label, you don't find out until months later when they're in customer hands and they start complaining, and now you have to unwind your manufacturing and fulfillment pipeline to get back all the devices you've shipped.
All that to protect against what attack? One where there's malicious software on the user's network that changes the device password before the user can? In that case, the user would just not use the camera because they can't access the feed.
> I agree that would be nice, but it also doesn't sound all that practical for a small vendor.
Personalizing / customizing per device always introduces a huge amount of complexity (and thus cost). However, this is TP-Link we're talking about, who definitely have the ability to personalize credentials at scale on other product lines.
And again, to be clear, I'm not trying to argue that the current way is some horrible disaster from TP-Link, just advocating for a better solution where possible. I think the current system reads as fine, honestly, it sounds like typical cobbled together hardware vendor junk that probably has some huge amount of "real" vulnerability in it too, but this particular bit of the architecture doesn't offend me badly.
> now you have to unwind your manufacturing and fulfillment pipeline to get back all the devices you've shipped.
This can be avoided with some other type of proof-of-presence side channel which doesn't rely on manufacturing personalization - for example, a physical side-channel like "hold button to enable some PKI-based backup pairing or firmware update mode." For a camera, there should probably be an option to make this go away once provisioning is successful, since you don't want an attacker performing an evil maid attack on the device, but for pre-provisioning, it's a good option.
> Personalizing / customizing per device always introduces a huge amount of complexity (and thus cost)
For a hardware product mass produced like this, they should already have a custom label that has the unique serial number on it which is also programmed into each device, so they should already have the infrastructure to do that (potentially as part of automated board testing/flashing).
Adding a randomly generated password is hardly more work once you have the ability to do that.
For a large manufacturer, it can be flashed onto the device automatically by a machine as part of the production line. That's not easy to set up, but basically something they already need to have set up (you don't want to have humans have to plug in boards to flash firmware and load serial numbers, etc. on to every unit).
> The device should ideally have some kind of secret material derived per device, like a passphrase generated from an MCU serial number or provisioned into EEPROM and printed on a label on the device.
It is better than simple secret like 12345678 but it can go wrong too, like in the case of UPC UBEE routers where the list of potential passwords can be narrowed down to like ~60 possibilities using a googled generator [1] whilst knowing only the SSID.
It did require firmware reverse engineering to figure out [2][3] but applies to most devices I've encountered. User should ideally always change the default password regardless.
AT&T routers, for example, ship like this. There's a wifi network and a wifi password printed onto the device.
But that also means then that often anyone with physical access can easily get into the device. The complicated password provides an additional layer of illusion of security, because people then figure "it's not a default admin password, it should be good". The fundamental problem seems to be "many people are bad at passwords and onboarding flows", and so trying variations on shipping passwords seem to result in mostly the same problems.
That's fair, though at least resetting would indicate that an attack happened. Default passwords and printed passwords can result in undetected attacks, which are arguably worse.
I feel seen. Why is the security illusory? I still don't understand the problem with this. Is the concern that someone will break into my house to covertly get access to my wifi password?
These may be illegal in some jurisdictions due to accessibility laws, and are a bad idea in general, for these reasons as well as unattended configuration scenarios.
If you buy the camera, plug it in, and forget to set it up, you just flat out can't use it right? I agree that proof of presence is way better but how many people are seriously going to be affected?
No, if you buy the camera, plug it in, and forget to set it up, then someone can use the default password and key material stored in the app to pretend to be the app and provision it on your behalf.
That's the only real vulnerability here, and it's no big deal, but it is A Thing and there is definitely a better way to do this that doesn't lose the freedom of full-offline.
There could be another scenario. I assume that factory resetting the camera will bring back the default password. Factory reset is a long press on the power button (in some cheap TPlink camera), so in theory someone with physical access can take control over the camera without you noticing until you try and use it.
I mean, given that it's updated after setup with the normal flow, I'm okay with it.
The thing I've most been convinced of in the past 5 years of building as much 'iot/smart home' stuff out as possible in my house is that nearly every vendor is selling crap that has marginal usefulness outside of a 'party trick' in isolation. Building out a whole smart home setup is frustrating unless it's all from one vendor, but there isn't one vendor which does all of it well for every need.
On my phone I have apps for: Ecobee, Lutron, Hue, 4 separate camera vendors[1], Meross, and Smart Life. Probably a couple more that I'm forgetting.
Only Lutron and Hue are reasonable in that they allow pretty comprehensive control to be done by a hub or HomeKit so I never have to use those apps.
It's been years since Matter and Thread were supposedly settled upon as the new standards for control and networking, but the market is, instead of being full of compatible devices, instead absolutely packed with cheap wi-fi devices, each of which is cloud-dependent and demands to be administered and even used day-to-day only through a pile-of-garbage mobile app whose main purpose is to upsell you on some cloud services.
[1] I admit the fact I have 4 is my fault for opportunistically buying cameras that were cheap rather than at least sticking with one vendor. But many people have a good excuse, perhaps one vendor makes the best doorbell camera, while another might make a better PTZ indoor camera.
> Even if your hardware doesn't support local APIs, there's a good chance someone has made an HA integration to talk to their cloud API.
And if they haven’t, you can pretty trivially write your own and distribute it through HACS (I’ve got three integrations in HACS and one in mainline now)
I love it! But my setup has a lot of sharp edges. It's a combo of things where the "standards compatible" way to connect to HA lacks things like camera control, by dastardly vendors like Chamberlain who basically killed HA support for spite, and finally, by having to use Google or Amazon for voice assistants.
My #1 wish would be for someone to build a HA-native voice assistant speaker. I'd pay $100 each for a smart speaker of the physical quality of the $30 Google Home Mini but which integrated directly with HA and used a modern LLM to decide what the user's intent was, instead of the Google Assistant or Siri nonsense which is like playing a text adventure whose preferred syntax changes hourly. I'd pay that plus a monthly fee to have that exist and just work.
Chamberlain can't change MyQ to get around the fact that HA can operate the switch in your garage with a simple controller attached to it. It is very annoying that they are anti-hacker though.
This M5 Stack ASR unit costs $7.50, and has a vocab of about 40-70 words. That's enough to turn on/off lights and timers. You might need to come up with your own command language, but all of the ASR is extremely local
That is probably a great and fun way to solve the problem for those with even a little free time.
Sadly for family reasons I sadly can't take on projects that require more than a few minutes, so I'm holding out hope for someone to bridge the gap between the "project boards that require writing a bunch of code to interface with Home Assistant and define all of its possible abilities and commands" and "dumb as a post Google thing that you just plug in" with a hardware device that is easy to connect to HA and starts out doing what the Google thing can do, but smart instead of stupid like the legacy voice assistants are.
I don't think we will ever understand what you saw if we continue to cling to our view of what reality actually is. I think it's unlikely whatever it is originated in some distant galaxy far far away (aliens, little-green men, spacecraft, etc). I think it's something fundamental and it's here on this planet and I don't think that sits well with people. Like many things, we have to look within rather than what's out there.
Do you think WSJ is doing this deliberately? Leaving out the part about how it's actually humans behind it. I've heard of this a long time ago, but I see this is a new story. Probably a hundred generations of monkeys have learned this behavior and they don't need any training or handling at this point!
