kgeist

43 ms·

691.
by kgeist 3y ago
I understand that it's a very interesting engineering problem - to create distrubuted coroutines and make them work in Python, but I'm not sure I understand why I would want to use this concept in our projects. I understand the id
692.
by kgeist 3y ago
In our code, the only DB-related piece of code which is aware of multi-tenant databases is the getDBConnection(accountId) function. Once you have the connection, you execute exact same SQL queries as before. The function is hidden deep insi
693.
by kgeist 3y ago
What do you mean by "rewriting everything"? Or maybe your definition of "per-tenant database" is different from mine. In our product, it's just a small layer which routes requests to the target organization's D
694.
by kgeist 3y ago
Or "delete all your comments" as a user message on a forum.
695.
by kgeist 3y ago
>Most engineering teams have some sort of approval process, often called a technical review, which is usually reserved for the later stages in a project. The problem with late-stage technical reviews is that when they happen too late, li
696.
by kgeist 3y ago
The site lists all those features but it's not clear what problem it's trying to solve. Only in the FAQ at the end of the page it's said it's a service mesh basically with added bells & whistles, from what I understa
697.
by kgeist 3y ago
Tried it with the local alternative to Google Disk. Oh my... Immediately found lots of private data, including photos of credit cars (with security codes), scans of IDs, passports... How do you report a site?
698.
by kgeist 3y ago
I wonder what are the security implications. I just opened a random URL and an arbitrary application was launched inside my Internet OS desktop. What if someone gives me a link to a malicious password stealer, for example? I think a permiss
699.
by kgeist 3y ago
ChatGPT was additionally finetuned after the initial training of the base model, so it's not completely clear if it truly represents the original distribution anymore. I use open weights models and they repeat such phrases less frequen
700.
by kgeist 3y ago
ChatGPT loves to add "it's important to remember" to every second answer. Also "multifaceted" and a few other adjectives like that.
701.
by kgeist 3y ago
Interesting that the article itself gives AI-generated vibes: >It's important to remember that while Wikipedia is "The Free Encyclopedia that Anyone Can Edit,"
702.
by kgeist 3y ago
Russia's prescriptions, patient history etc. all still happen on paper. The Kremlin uses typewriters for sensitive data [0] etc. So there's often nothing to hack. [0] https://www.npr.org/sections/thetwo-way&#x
703.
by kgeist 3y ago
The article says: >In Tibet, Dr. Huerta-Sánchez and her colleagues have found a Denisovan gene that helps people survive at high altitudes Can't it be simply convergent evolution?
704.
by kgeist 3y ago
>Code that isn't tested, isn't done. Tests not only verify the expectations, but also prevent future regression A few weeks ago someone at the company implemented a feature without tests. It worked perfectly, everyone was happy
705.
by kgeist 3y ago
The main hurdle is lack of multilingual support in open weights models.
706.
by kgeist 3y ago
Are submissioms manually reviewed by a human? Otherwise it can be abused for spam. Is it reindexed from time to time? Say, the submission looks okay, and then they change the content to something spammy. Maybe a Report button could be added
707.
by kgeist 3y ago
So the questions should come before the content and it might work? I think that's how also RWKV works.
708.
by kgeist 3y ago
If Mamba selectively forgets "unnecessary" details, can it repeat the input verbatim (if asked)?
709.
by kgeist 3y ago
It's verbose because it was finetuned to be "helpful" (the way OpenAI sees it). You can fix it with a system prompt, or finetune the base model with the format you want. Same with grounding it with RAG. Sure if you take a van
710.
by kgeist 3y ago
The problem is probably specific to the default system prompt in the default ChatGPT UI version. An open weights LLM can be made into a pretty engaging bot with the right prompt/model/finetuning. I use an uncensored LLM with a sp
711.
by kgeist 3y ago
It can easily be made less verbose with the right prompt.
712.
by kgeist 3y ago
I think DB engines should simply have some sort of a default option in the interactive mode where if you write "delete from table", it asks "Are you sure? You're going to wipe out the entire table! Y/N". Would&
713.
by kgeist 3y ago
At our company, we have "an immutable DB", too, but when there's a critical emergency (say, full downtime), we can apply fixes manually. In that case, we run the tests after applying the fix.
714.
by kgeist 3y ago
Not to say, miscommunication/misunderstanding happens between people all the time. And software has already killed/injured people without any AI involved. Nothing specific to AI here. It's more about a lack of adequate QA.
715.
by kgeist 3y ago
An update on my endeavour: so, model switching is very costly under llama.cpp (I have to switch between Llama and Phi2 because my GPU has low amounts of VRAM). And this switch (reloading the weights into VRAM) defeats the whole purpose of t
716.
by kgeist 3y ago
>this incident shows that even a minor mistake from the AI builders could lead to the AI freeing itself from our control. There's most likely a hidden prompt which tells Gemini to make people racially diverse. And Gemini complied pe
717.
by kgeist 3y ago
We have a script which recreates the dev DB, but developers must run the command manually so that there was an understanding what's happening.
718.
by kgeist 3y ago
Reminds of this classic thread: https://www.reddit.com/r/cscareerquestions/s/cqrama0L1z >accidentally destroyed production database on first day of job Same symptoms: while developing a feature locally, th
719.
by kgeist 3y ago
What prompts/settings do you use for Phi-2? I found it completely unusable for my cases. It fails to follow basic instructions (I tried several instruction-following finetunes as well, in addition to the base model), and it's been
720.
by kgeist 3y ago
I have a locally hosted chatbot based on Llama which is always antagonist (the model is uncensored + there's a system prompt to guide it). There's an example floating around on the internet: if you say "5+2=8" to ChatGPT

More ›