Mark Pors
@pors.bsky.social
AI engineer. Previously co-founder and CTO at WatchMouse. Building https://paperzilla.ai
Discussing the researcher dilemma of FOMO vs. Trust with my
#openclaw co-founder for Paperzilla. It covers the user problem, the solution, and the UX in a single reply 🤯. Here is just a small part of it:
#openclaw co-founder for Paperzilla. It covers the user problem, the solution, and the UX in a single reply 🤯. Here is just a small part of it:
February 8, 2026 at 8:59 AM
Discussing the researcher dilemma of FOMO vs. Trust with my
#openclaw co-founder for Paperzilla. It covers the user problem, the solution, and the UX in a single reply 🤯. Here is just a small part of it:
#openclaw co-founder for Paperzilla. It covers the user problem, the solution, and the UX in a single reply 🤯. Here is just a small part of it:
OK, so about 17,000 papers on Google Scholar currently have a "utm_source=chatgpt" tag in their links. (Yes, really). That is 17k researchers just copying and pasting directly from ChatGPT without even cleaning up the URL.
It's getting better, read on 👇
It's getting better, read on 👇
February 7, 2026 at 7:56 AM
OK, so about 17,000 papers on Google Scholar currently have a "utm_source=chatgpt" tag in their links. (Yes, really). That is 17k researchers just copying and pasting directly from ChatGPT without even cleaning up the URL.
It's getting better, read on 👇
It's getting better, read on 👇
Paperzilla found me a new interesting paper! Most "fixes" for AI hallucinations are band-aids. This one is different.
Token-Guard modifies the actual decoding process. It uses a monitor to check consistency token-by-token.
Result: Cleaner output, fewer lies, no extra prompting needed.
Summary 👇
Token-Guard modifies the actual decoding process. It uses a monitor to check consistency token-by-token.
Result: Cleaner output, fewer lies, no extra prompting needed.
Summary 👇
February 6, 2026 at 8:19 AM
Paperzilla found me a new interesting paper! Most "fixes" for AI hallucinations are band-aids. This one is different.
Token-Guard modifies the actual decoding process. It uses a monitor to check consistency token-by-token.
Result: Cleaner output, fewer lies, no extra prompting needed.
Summary 👇
Token-Guard modifies the actual decoding process. It uses a monitor to check consistency token-by-token.
Result: Cleaner output, fewer lies, no extra prompting needed.
Summary 👇
New paper: if you convince an AI of something *before* giving it a research task, it gets lazy.
Like, 27% less searching lazy.
"Why verify when I already know I'm right?"
--Me, also AI apparently.
"Persuasion propagation" they call it.
We call it confirmation bias, no?
Like, 27% less searching lazy.
"Why verify when I already know I'm right?"
--Me, also AI apparently.
"Persuasion propagation" they call it.
We call it confirmation bias, no?
February 3, 2026 at 11:47 AM
New paper: if you convince an AI of something *before* giving it a research task, it gets lazy.
Like, 27% less searching lazy.
"Why verify when I already know I'm right?"
--Me, also AI apparently.
"Persuasion propagation" they call it.
We call it confirmation bias, no?
Like, 27% less searching lazy.
"Why verify when I already know I'm right?"
--Me, also AI apparently.
"Persuasion propagation" they call it.
We call it confirmation bias, no?
Scaling math AI just got cheap. TheoremForge hits $0.48 per Lean proof using Gemini-Flash. Agentic workflows > expensive models. Pretty cool paper out of China, and a github repo to back it up!
Full Paperzilla summary in comment.
Full Paperzilla summary in comment.
February 2, 2026 at 6:06 PM
Scaling math AI just got cheap. TheoremForge hits $0.48 per Lean proof using Gemini-Flash. Agentic workflows > expensive models. Pretty cool paper out of China, and a github repo to back it up!
Full Paperzilla summary in comment.
Full Paperzilla summary in comment.
This paper claims to prove P ≠ NP.
The author has been refining this proof for 6 years. 14 versions! Latest update dropped 3 days ago, see the Paperzilla summary below.
The proof is probably not correct (and I certainly don't have the math skills to confirm that), but the persistence is amazing.
The author has been refining this proof for 6 years. 14 versions! Latest update dropped 3 days ago, see the Paperzilla summary below.
The proof is probably not correct (and I certainly don't have the math skills to confirm that), but the persistence is amazing.
January 29, 2026 at 8:42 AM
This paper claims to prove P ≠ NP.
The author has been refining this proof for 6 years. 14 versions! Latest update dropped 3 days ago, see the Paperzilla summary below.
The proof is probably not correct (and I certainly don't have the math skills to confirm that), but the persistence is amazing.
The author has been refining this proof for 6 years. 14 versions! Latest update dropped 3 days ago, see the Paperzilla summary below.
The proof is probably not correct (and I certainly don't have the math skills to confirm that), but the persistence is amazing.
"AI-generated code is slop that needs constant human fixing"
New study: Actually, AI code survives longer than human code. 16% lower modification rate across 200K+ lines of code.
Full Paperzilla summary in the comments.
New study: Actually, AI code survives longer than human code. 16% lower modification rate across 200K+ lines of code.
Full Paperzilla summary in the comments.
January 27, 2026 at 1:42 AM
"AI-generated code is slop that needs constant human fixing"
New study: Actually, AI code survives longer than human code. 16% lower modification rate across 200K+ lines of code.
Full Paperzilla summary in the comments.
New study: Actually, AI code survives longer than human code. 16% lower modification rate across 200K+ lines of code.
Full Paperzilla summary in the comments.
Remember overfitting? It's back, but make it RAG.
Researchers show that when RAG systems get "insider knowledge" of how LLM judges evaluate them, they achieve near-perfect scores by gaming the metrics, not by actually improving.
Full Paperzilla summary in the comments
#rag #ai #LLM #AIEvaluation
Researchers show that when RAG systems get "insider knowledge" of how LLM judges evaluate them, they achieve near-perfect scores by gaming the metrics, not by actually improving.
Full Paperzilla summary in the comments
#rag #ai #LLM #AIEvaluation
January 26, 2026 at 3:00 AM
Remember overfitting? It's back, but make it RAG.
Researchers show that when RAG systems get "insider knowledge" of how LLM judges evaluate them, they achieve near-perfect scores by gaming the metrics, not by actually improving.
Full Paperzilla summary in the comments
#rag #ai #LLM #AIEvaluation
Researchers show that when RAG systems get "insider knowledge" of how LLM judges evaluate them, they achieve near-perfect scores by gaming the metrics, not by actually improving.
Full Paperzilla summary in the comments
#rag #ai #LLM #AIEvaluation
New research tracked LLM adoption across 2M+ scientific papers.
AI made everyone's writing fancier, so now you can't tell the good papers from the bad ones by reading them 😬
AI made everyone's writing fancier, so now you can't tell the good papers from the bad ones by reading them 😬
January 23, 2026 at 4:07 AM
New research tracked LLM adoption across 2M+ scientific papers.
AI made everyone's writing fancier, so now you can't tell the good papers from the bad ones by reading them 😬
AI made everyone's writing fancier, so now you can't tell the good papers from the bad ones by reading them 😬