It’s been insisting for an hour that Lindsey Graham is alive, all the news sources contradicting him it are satirical or spoofed, and part of his “evidence” are no releases from the governor’s office, and articles are dated in the future. 🤣🤣🤣
i keep sending deepseek quotes from various popular media to see if it can figure out where they come from and i’ve found it has a 50% chance of attributing the quote to not just the wrong characters entirely but to entirely different works in entirely different mediums in entirely different genres
Yes, it’s tragicomic to me that techbros think this is the future of governance.
You don’t have “web search” highlighted so it’s just going off of its long-term memories back when it was originally trained. If you enable “web search” you won’t have this problem. Even in my local instance of Qwen3.6-27B it answers correctly, because I run it through the Hermes client which can use a web search back-end.
──────────────────────────────────────── ● Is Lindsey Graham still alive? Initializing agent... ──────────────────────────────────────── ┊ 🔍 preparing web_search… ┊ 🔍 search Lindsey Graham 2026 current status 1.0s ╭─ ⚕ Hermes ───────────────────────────────────────────────────────────────────────────────────────────╮ I'm seeing some search results that reference Lindsey Graham's death, but these need to be verified against reliable sources before drawing any conclusions. Let me dig deeper. ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ ┊ 🔍 preparing web_search… ┊ 🔍 search Lindsey Graham death 2.2s ╭─ ⚕ Hermes ───────────────────────────────────────────────────────────────────────────────────────────╮ Let me verify this by checking one of these sources directly. ╰──────────────────────────────────────────────────────────────────────────────────────────────────────╯ ┊ 📄 preparing web_extract… ┊ 📄 fetch www.cnn.com 12.4s ╭─ ⚕ Hermes ───────────────────────────────────────────────────────────────────────────────────────────╮ No, Lindsey Graham is no longer alive. According to multiple credible news sources (CNN, CBS News, ABC News), Senator Lindsey Graham died on July 12, 2026, at age 71. According to the CNN report: - He died after a "brief and sudden illness," following an aortic dissection (a tear in the inner layer of the aorta) caused by arteriosclerotic cardiovascular disease, per the DC medical examiner's preliminary finding. - His death occurred shortly after he returned from a trip to Ukraine. - He had spoken with President Trump just hours before his death. A special primary election is currently underway in South Carolina to determine his replacement in the U.S. Senate. Republican candidates Darline Graham (his adopted sister) and Ralph Norman competed in a primary runoff on August 25, 2026, with Darline Graham winning the GOP nomination.There has basically been zero progress against fixing hallucinations in the AI industry. All top AI models still hallucinate badly. The way the industry has instead been handling it is to just give AI models better context that contains the right information to begin with so they’re less likely to hallucinate. That’s why web search is added, if you enable web search it will pull relevant data in from the web into its context window which will make it more likely to give you the correct answer.
It’s also why the industry has been moving away from chatbot AI to agentic AI. Agentic AI can have a much more complicated loop of actually verifying its own claims before giving them to you. They still hallucinate all the time, but they can catch their own hallucinations and correct for them before actually finishing the loop, so in practice you stop noticing the hallucinations even though they are still there.
It’s ultimately an alignment problem. The AI model is ultimately trained to provide answers which please a human reviewer. If it simply does not have the answer, but judges that bullshitting an answer will please the human reviewer more than just saying it doesn’t know, it will bullshit an answer. No one has figured out how to truly fix this alignment problem at its root.
So, instead, the industry has moved towards just making sure the AI has the right answer to begin with, giving it the proper tools to be able to search for it on its own, find the answer, and even verify and double-check itself it pulled the correct thing, before reporting it. If you have these tools disabled then it’s going to hallucinate bad.
Is there a way to enable web search when the LLM is accessed from the internet?
If you’re using the browser chat you just have to click the little globe icon and make sure it is highlighted. That globe icon means web search.
There is no such icon, and I find it in no menu. I’m not saying it’s not there, my vision isn’t optimal, despite corrective lenses. Am I missing it?
llms don’t make good search engines… they’re very susceptible to returning garbage results, and even misinterpreting results.
The disclaimer at the bottom of the page is the cherry on top: “AI responses may be inaccurate—review them before you rely on them.” If so, then why the f would I need to use AI?
To find obscure information or sources. I’ve been messing with Qwen for a couple of months, and the biggest limitations to date have been the hosting state: the two I have bookmarked are hosted in Singapore (in the screenshot), and Germany
(I didn’t test to see if the German hosted Qwen answered differently),German hosted Qwen gives the correct date, but doesn’t recognize Darline Graham; and properly structuring my own queries, whether for text or image output.The instance in Singapore discusses communist theory and religion more freely. On either site, Qwen has taken to simply claiming I’ve used my token allotment, rather than telling me directly it can not discuss sensitive topics like it did in early trials, and both tend to cite official Western narratives. The German hosted Qwen will shut down the season if you push back too hard; the Singaporean gives a little more leeway if you properly structure queries, e.g., more inquisitive than accusatory.
I queried about the book, Operation Gladio, on the Singaporean Qwen, and it told me the book was sensationalist, with dubious sourcing (fair enough). When I pushed by mentioning Operation Condor, it recommenced J. Patrice McSherry Predatory States: Operation Condor and Covert War in Latin America or Daniele Ganser NATO’s Secret Armies, instead.
A pity I’ve no way of registering for Chinese hosted Qwen or Deepseek, without a Chinese telephone number, or means to store and host them, myself. I don’t really understand why I can have a Rednote account, but can’t use a Chinese hosted AI.
Anyway, checking sources is best practice, even if I sometimes neglect to do.
In the I was able to sign up and buy DS tokens for vibecoding projects earlier this year, though I used the Chinese platform landing page (can use translate, all I can read is numbers and things like pay and download) and my Google account to do so. If I’m going to get in trouble with the law its going to be for existing while minority like five times, long before using a foreign LLM gets me.
I also use DS’ app on my phone and have been for 2yrs or so since its snappy on my now 4yro device, though I really like Z.ai since that’s no log in and browser accessible, but it seems like it gets busy often.
It’s Qwen3.6, but there’s a 1G version that can be run on a phone: https://prismml.com/news/ternary-bonsai
Hmm. Thank you. I’ve got to get to solving legacy Android device issues, too.
Like the other comment, obscure information. Also the consolidation of information (again, like always, as long as you verify it). I use Debian 13 and a bunch of information is either; huge amount of jargon, sparse, or dispersed unnecessarily. For instance, the default Debian graphic drivers for Nvidia cards are outdated to the point that my computer was becoming finicky with certain games and programs. I could not find out any good and concise information about solving it through using the official nvidia database, alot of it was either contradictory, jargon, or just shit. Using Qwen helped me out big time surprisingly. I also use it to make my Stellaris mod lol. Like any personal tool, it should reduce the amount of time used to commit certain work. I use it as information gathering rather than “make this for me”.
Oh yes, it’s very good at explaining technical things (medical, IT, religious doctrines (it gave me very good, simple differences/similarities in various sects of Islam and Buddhism, compared and contrasted Dzogchen and Gnosticism which matches my own understanding, Mahayana and Theravada, political doctrines with similarities and differences that match what I think I understand, but I have to read up on Singapore, it claimed it’s a technocracy, and I don’t really know).
We also discussed whether or not the US would be better off Balkanized (Qwen thinks so, with the exception of poors in the Deep South, of which he paints an extremely bleak picture, with which I happen to agree).
Qwen thinks there are protections for the USD despite selloff of bonds and USA printing money, and that Western states with their economies tied to the USD will certainly see a “deep recession,” but not a depression. I’m very poor on grasping economics, so I can’t say whether I agree or not.
People are dangerously stupid and they’re taking everyone down with them.
