Trey Menefee (🇭🇰 🇹🇭 🇺🇸)
LightNews LightNews
banner
comparativist.bsky.social
Trey Menefee (🇭🇰 🇹🇭 🇺🇸)
@comparativist.bsky.social
540 followers 2.3K following 150 posts
Founder of Synthetic Intelligence Group
🏝️ Escaped Florida Man | South China (20 yrs)
🇭🇰 Hong Konger | 🇹🇭 In Thailand |
Posts Replies Media Videos
comparativist.bsky.social
Perplexity FTW
January 4, 2025 at 2:02 PM
comparativist.bsky.social
Sure
The hallucinations you’re encountering occur because you’re asking the model to infer knowledge it doesn’t possess, which it compensates for by generating plausible but inaccurate outputs due to its design as a probabilistic next-token predictor, not a database. To improve prompts, understand the model’s strengths (broad generalizations, well-established facts) and weaknesses (specific datasets, niche queries) and recognize that reinforcement learning biases it toward being “over-eager to please.” For fact-checking, tools like Perplexity or models integrating chain-of-thought reasoning and web grounding are better equipped to handle such queries, while current models like ChatGPT-4o are limited by their inability to reason robustly or validate niche, underrepresented datasets.
January 4, 2025 at 1:30 PM
comparativist.bsky.social
Jumped on a 15¢ train to Samut Sakhon today
January 2, 2025 at 7:07 PM
comparativist.bsky.social
long-form response
I would push back, respectfully, on the premise that these models need to meet the deterministic standards of traditional software systems. These are fundamentally different tools, operating probabilistically, and their purpose is not to replicate the reliability of a calculator or spreadsheet. Instead, we should be focusing on informed use—identifying where these models excel and validating best practices for those tasks. For instance, LLMs like GPT-4 are remarkably effective at summarizing dense texts, performing sentiment analysis, organizing data into structured formats, and generating coherent prose for ideation. These are not areas where they just compete with humans; they often exceed human capability. But when it comes to more technical tasks—like math or running simulations—current use demands a much more cautious, evidence-based approach. That said, I think we share common ground on the need for transparency, both from vendors and users. If someone told me they used ChatGPT to run a Monte Carlo simulation, I’d have to question whether it truly happened, because OpenAI, for example, doesn’t provide sufficient documentation about what’s happening under the hood. We don’t know the size of the context windows, we don’t know the full extent of model updates, and we don’t know when it’s simulating versus genuinely executing code. Researchers, in turn, need to clearly describe how they are using these tools in their methodology sections, so their claims can be interrogated and replicated. This isn’t about meeting outdated standards but about adapting to the specific affordances and limitations of these systems while building trust through clarity and rigor.
January 2, 2025 at 6:04 PM
comparativist.bsky.social
Happy New Year from BKK!
December 31, 2024 at 5:14 PM
comparativist.bsky.social
Final sunset of 2024 #bkk
December 31, 2024 at 11:40 AM
comparativist.bsky.social
Bang Khun Thian (Bangkok) two days ago
December 23, 2024 at 11:02 AM
comparativist.bsky.social
a non-LOTR character who resisted the One Ring IRL
December 23, 2024 at 10:08 AM
comparativist.bsky.social
Check out my new shirt
December 23, 2024 at 2:56 AM
comparativist.bsky.social
Vietnam karks are quite mid

El Nido is the best and part of same system that runs through Guilin too
December 22, 2024 at 4:15 PM
comparativist.bsky.social
Picked this up for $5 today. Maybe IR photography.
December 22, 2024 at 3:52 PM
comparativist.bsky.social
Packard Bell i486
December 22, 2024 at 3:12 PM
comparativist.bsky.social
This is my barbers band music.youtube.com/playlist?lis...
December 20, 2024 at 1:35 PM
comparativist.bsky.social
Cool
December 11, 2024 at 6:44 AM
comparativist.bsky.social
27km away
November 30, 2024 at 1:38 PM
comparativist.bsky.social
November 27, 2024 at 6:58 PM
comparativist.bsky.social
if you see this…
November 26, 2024 at 1:30 PM
comparativist.bsky.social
Just do Cantonese roast duck or bbq pork ffs
November 25, 2024 at 1:51 PM
comparativist.bsky.social
I just invented red curry noodle-rice.

Please alert the relevant Nobel Committee. I will not be taking questions at this time.
November 25, 2024 at 10:51 AM
comparativist.bsky.social
dinner #DiningWithTrey #BKK #Thailand
November 23, 2024 at 11:53 AM
comparativist.bsky.social
Absolutely gorgeous day in BKK
November 23, 2024 at 5:33 AM
comparativist.bsky.social
Three days later
November 21, 2024 at 1:02 PM
comparativist.bsky.social
ok. @armscontrolwonk.bsky.social
November 21, 2024 at 12:55 PM
comparativist.bsky.social
jfc #ukraine
November 21, 2024 at 12:47 PM
comparativist.bsky.social
HOLY SHIT those are MIRVS coming down!!!!!
November 21, 2024 at 12:45 PM