renonce

Just make the RAM mismatch a warning that voids warranty rather than a hard error. This is no excuse for forcing hardware pairing

I built a project achieving similar goals. You launch a web server then connect to it using either browser or Android app, then create a session to talk to Claude Code. The sessions are synchronized in real time across all devices and automatically saved to disk and continued when server restarts. Recently I've added features to schedule tasks in the future and to assemble agent teams. The project is mostly vibe-coded with Opus 4.6 with few supervision beyond trying its functionalities out.

Project is at http://github.com/vincent-163/claude-code-multi/. Can be installed easily with nodejs.

Please provide feedbacks and suggestions!

My experience is that experts are not separated in any intuitive way. I would be very interested (and surprised) if someone manages to prune a majority of experts in a way that preserves model capabilities in a specific domain but not others.

See https://github.com/peteryuqin/Kimi-K2-Mini, a project that keeps a small portion of experts and layers and keep the model capabilities across multiple domains.

In Chrome you can F12 and go to "Network" tab and then refresh the page. Choose the first file in the list (that's the HTML itself) and you will find "Response Headers" in the "Headers" panel, which includes Last-Modified. It's a bit deep, which makes sense as it's rarely useful.

> That warning showed the patient's white blood cell count was "really, really high," recalled Bell, the clinical nurse educator for the hospital's general medicine program.

I’m not sure how an alarm for “high white cell count” should have had so much impact. Here in China once the doctor prescribes a finger blood test, we sample finger blood after lining up for 15 minutes, and the result is available within 30 minutes. The patient prints the results from a kiosk and any patient who cares enough about their own health will see the exceptionally high white cell count and request an urgent appointment with the doctor for diagnosis right away. Even in normal cases we usually have the doctor see the report within two hours. Why wait several hours?

> While the nursing team usually checked blood work around noon, the technology flagged incoming results several hours beforehand.

> But in health care, he stressed, these tools have immense potential to combat the staff shortages plaguing Canada's health-care system by supplementing traditional bedside care.

This sounds like the deaths prevented by this tech are caused by delays and staff shortage and what this tech does is to prioritize patients with serious issues? While I appreciate using new tools to cut deaths, it looks like the elephant in the room is staff shortage?

> Also, there is no way to know which related site the user is logged in to, so they would have to prompt for every one of their sites.

This is not how it works. The mechanism is about allowing a cluster of websites to choose a single first party domain and have all of them share cookies together, not sharing arbitrary cookie from arbitrary domain, otherwise it would create loopholes in connected components that bring back the downsides of third-party cookies. What you mentioned should be done using SSO.

After thinking about it a bit more, I have a clearer picture of how it should work in my mind:

* All cookies are double-keyed: the primary key is the origin of the top-level page and the secondary key is the origin of the page that sets the cookie, just like how partitioned cookies work right now.

* stackoverflow.com uses a header, meta tag or script to request changing its primary key domain to “stackexchange.com”

* The browser makes a request to https://stackexchange.com/domains.txt and make sure that “stackoverflow.com” is in the list, authorising this first-party domain change

* When the user agrees to the change, the page is reloaded with stackexchange.com as the primary key, thus stackoverflow.com can obtain login details from stackexchange.com via CORS or cross site cookies.

* A side effect is that all cookies and state are lost when switching the first-party domain. Should stackoverflow.com be acquired by a new owner, say x.com and changes its first-party domain to x.com, all cookies on stackoverflow.com are lost and the user will have to login on x.com again, maybe using credentials from stackexchange.com. It’s unfortunate but it works around the issues mentioned in the post in a clean way, avoiding loopholes that transfer cookies by switching the first-party domain frequently.

> I can't log in to stackoverflow.com, then go to superuser.com and already be logged in.

I would expect a popup like “This site wants to share cookies with stackexchange.com, press Allow to sign in, press Reject to reject forever or press Ignore to decide later”. Takes a single click to enjoy the benefits of both worlds. The mechanism should make sure that every website has a single “first-party domain” shared across all subsites and that first-party domain must not share cookies with any other site than itself to minimize confusion.

Looks like a perfect use case for LLM: generate that JSON-LD metadata from HTML via LLM, either by the website owner or by the crawler. If crawlers, website owners doesn’t need to do anything to enter Semantic Web and crawlers specify their own metadata format they want to extract. This promises an appealing future of Web 3.0, not by crypto, defined not by metadata but by LLMs.

> The settlement, announced Tuesday, does not act as an admission of guilt and Meta maintains no wrongdoing.

> In 2011, Meta introduced a feature known as Tag Suggestions to make it easier for users to tag people in their photos. According to Paxton’s office, the feature was turned on by default and ran facial recognition on users’ photos, automatically capturing data protected by the 2009 law. That system was discontinued in 2021, with Meta saying it deleted over 1 billion people’s individual facial recognition data.

> The 2022 lawsuit

> We are pleased to resolve this matter, and look forward to exploring future opportunities to deepen our business investments in Texas, including potentially developing data centers

Each statement makes it increasingly harder to view it as a fine than a tax. An offence that lasted 11 years and got prosecuted a year after it ended can be explained in no other way than being an excuse dug out of the ground to make a ransom

> This whole HN post could just be bots all the way down and it'd still be an interesting read through the comments.

The assumption is that the comments are a function of the post and the present public info. Real world comments can disclose private information (My Google account banned), make real impact (S*e/C*e support site), connect to celebrities (I'm Karpathy ask me anything). And even if the assumption holds and the site is a sample from a probability distribution, the particular sample can be referenced in other sites so it makes sense to check what everyone is viewing right now.

You can usually flash BIOS while wiping your computer in the same way that a malware does except in very rare cases. Also Secure Boot doesn't remove the kind of rootkit that doesn't get removed along with the storage since it has to boot from your hard drive anyway.

I feel many security researchers like to overemphasize the importance of certain security practices (the most common one being "longer and random password with symbols and upper case letters") without considering its costs, trouble, and human's lazy nature. Forcing long passwords causes people to use repetitive or easy to remember words, enforcing Secure Boot doesn't work if it gets in the way of normal boots. Making sure that these security mechanisms "just work" is as important as enforcing rules like these.

A natural question is whether Secure Boot is the right place to protect against the type of attack mentioned in the post. Given that we've already invested a lot of effort in fixing kernel privilege escalations, and any program able to install BIOS rootkits can access all data and modify any program anyway, what justifies the extra complexity of Secure Boot (which includes all the extra design necessary to make it secure, such as OS'es robust to tampering even with kernel privileges)? I mean, why invest so much in Secure Boot when you could harden your kernel to prevent tampering BIOS in the first place?

It’s going to be significantly faster very soon, we have seen how AlphaGo evolved into KataGo which is many magnitudes more compute efficient

So if I understand correctly, this puts x.com under the same entity as twitter.com so third party cookies are allowed between x.com and other twitter sites?

This website is indeed hosted by Cloudflare.

Asking it “3.11和3.8哪个大” (meaning “Which one is larger, 3.11 or 3.8?” in Chinese) and it answers 3.11 more than half of the time. I assume it’s because Python 3.11 is larger than Python 3.8. While it does work in its native language English, this failure doesn’t give me much confidence in its reliability, as we don’t know why it works in one language but not the other yet.

Even as a self-motivated learner I fail to see the bigger impact of AI. For a “virtual Feynman” I would prefer the online video courses and books which exist without AI. The best I expect an AI to do is to answer my questions and confirm my understandings. At AI’s current state I can use it as a better search engine but due to hallucinations I can’t expect reliable answers yet.

My bank app asks for different tokens for different operations. A code for login, a code for transfers (the code needs to be generated with the payee account number as input). So it’s not a problem of tokens vs SMS.

Every VPN advertised in China takes DPI in account and there is a very huge community around how to circumvent GFW. The previously popular protocol was Shadowsocks (until GFW blocked fully random traffic) and now I use Hysteria2 (which masquerades itself as QUIC) and it has not been blocked yet.

In case your favorite VPN provider does not provide these protocols you can buy a VPS and deploy one of these protocols by yourself.

Not showing up is an arbitrage opportunity which means everyone can sue Facebook and get a win for free, which would not be sustainable.

I personally find such comparisons unfair. A good comparison should optimize for each device configuration, which means use a model within the VRAM limit and quantize to 8 bits where it boosts performance etc and avoid shortcomings of both devices unless necessary.

> 1.4. Mispairing database files and hot journals

I’m surprised it would corrupt the database. The database could have included a unique identifier as well as in the journal and refuse opening the database if the identifiers don’t match.

LLMs does have that, or at least it’s very likely that we will eventually be able to manipulate LLMs in a modular way (see https://news.ycombinator.com/item?id=40429540). One point remains: humans learn language with much fewer tokens than LLMs need, which suggests presence of a priori knowledge about the world. The LLM metaphor is finetuning, so babies are born with a base model and then finetuned with environment data, but it’s still within LLM scope.

You don’t want to be responsible for some random receipt 10 years ago that allegedly commits tax fraud, allow influencers to manipulate you into buying things or supporting campaigns that you immediately regret 1 minute later, receive lots of spams in your email and social accounts, etc, and the stakes get higher as you climb up the social ladder. Many of these don’t matter for people with little stakes as they have nothing to lose.

I mean I respect him for his contributions but what about being more specific with the contributions, like “I brought the idea” or “I supervised the research” or even “I managed the lab and arranged for the equipments with which my students conducted research”, any would be perfectly valid contributions to science. But it seems like the academic norm is just to boast “I was an author of these shiny new papers” and the details of how are obscured away almost intentionally.

I did read the original post in Chinese, though not this post in English. Unfortunately beyond the claim "search engines show up zero results" the other claims like 99.99% pages disappearing doesn't fully match my experience. I have lots of threads between 2000~2010 on Baidu Tieba (Chinese Reddit), and most of these threads representative of my memories from that decade are alive (I can even access them anonymously). The earliest video on bilibili.com still available is av7 (7th video since its creation). While some websites like RenRen did disappear together with its history, most websites that remain relevant didn't delete its history in its entirety (lots of content did disappear because of modern moderation applied retroactively to old content, but that's case by case).

Why expect search engines to return historical data accurately? Modern search engines have a lot of tasks like combating CEO and returning up-to-date data, and they have no incentive to preserve history as old as 2005 as it’s very likely that any page from that date has been superseded by more relevant articles. The task of preserving history should be delegated to archive.org and search engines are just not well positioned.

The closest research would be the Chinchilla scaling laws, which estimates the final loss as a function of number of parameters and tokens. Set the number of tokens to infinity would give a good estimate of minimum achievable loss.

For a 4096x4096 matrix its symmetry group has size 4096!. The Stirling’s approximation gives ln(4096!)=4096ln(4096)-4096 which is about 10~11 bits per 4096 numbers. This is less than 0.003 bit per parameter saved.

Can't speak for QAT as I haven't yet dived into that area. I've quickly skimmed the BitNet and BitNet 1.58 paper. I think achieving comparable performance with a Llama model with the same number of parameters is impressive but unfortunately it seems they didn't release the training code so I can only tell from their paper. Fortunately they did talk about training details in the BitNet paper (not in BitNet 1.58 so I assume they remain the same):

> Mixed precision training. While the weights and the activations are quantized to low precision, the gradients and the optimizer states are stored in high precision to ensure training stability and accuracy. Following the previous work [LSL+21], we maintain a latent weight in a high-precision format for the learnable parameters to accumulate the parameter updates. The latent weights are binarized on the fly during the forward pass and never used for the inference process.

In this case there are two areas to optimize for: training efficiency and inference efficiency.

If I understand correctly, it stores the weights, gradients and second-moment estimates in FP32 like every other mixed-precision training (the Gopher paper has details on why storing them in FP32 is important), and quantized weights are used in forward pass. What I'm not sure is whether latent weights are used in backward pass, and my instinction is that the "Straight-through estimator" requires high-precision latent weights so they may still be needed. Training FLOPS can be roughly estimated as 6 FLOP per parameter per token, where 2 is forward pass, 2 is gradient computation and 2 is gradient accumulation (see https://medium.com/@dzmitrybahdanau/the-flops-calculus-of-la...). If only forward pass is quantized, this means only 1/3 of all FLOPS are optimized (and even then it has to be accumulated in FP32). So I'm skeptical of the gains in training efficiency here, and I can't find the numbers (how much energy or how much time is used for training, compared to regular FP16 mixed precision training? The papers boast inference energy savings which makes me even more skeptical of training energy savings)

For quantization efficiency, while QAT can certainly avoid the quantization step, PTQ methods are very cheap (usually <24 hours on RTX 4090 for Llama-2-70b) so I consider the cost of the quantization step negligible. There is not much difference in inference efficiency gains as PTQ and QAT can quantize to the same format. For final accuracy, unfortunately there is a lack of comparison between QAT and PTQ of fp16 models, and PTQ has the advantage of not requiring access to the original dataset, so I think it's very hard to make a fair comparison here but it's also likely the only area where QAT has actual gains compared to best PTQ methods.