Nish Tahir
@nishtahir.com
Principal Engineer (AI Research). Anti-hype. My opinions are my own. I try not to be, but I can and will be wrong sometimes.
Blog: https://nishtahir.com
Mastodon: social.nishtahir.com/@nish
Blog: https://nishtahir.com
Mastodon: social.nishtahir.com/@nish
This is an interesting claim that's been made that still needs to be tested for legitimacy but it puts forward an interesting way to think of LLMs.
the-decoder.com/gpt-5-allege...
the-decoder.com/gpt-5-allege...
GPT-5 allegedly solves open math problem without human help
GPT-5 solves an open math problem for the first time, and the mathematician behind it shows exactly which line came from which AI. Does science really need this level of transparency?
the-decoder.com
December 25, 2025 at 12:38 AM
This is an interesting claim that's been made that still needs to be tested for legitimacy but it puts forward an interesting way to think of LLMs.
the-decoder.com/gpt-5-allege...
the-decoder.com/gpt-5-allege...
Looks like someone goofed 😂 #pokemontcgpocket
December 19, 2025 at 12:21 AM
Looks like someone goofed 😂 #pokemontcgpocket
Looks like GitHub will now be charging users who bring their own runners to subsidize users who use their hosted runners.
resources.github.com/actions/2026...
resources.github.com/actions/2026...
December 16, 2025 at 10:10 PM
Looks like GitHub will now be charging users who bring their own runners to subsidize users who use their hosted runners.
resources.github.com/actions/2026...
resources.github.com/actions/2026...
I'm realizing that there's going to be a lot more one-off notes in repos like those signs at bars that all have some bizarre stories attached to them.
December 15, 2025 at 12:03 AM
I'm realizing that there's going to be a lot more one-off notes in repos like those signs at bars that all have some bizarre stories attached to them.
Gemini's "You're absolutely right!" is "You have hit the nail on the head".
December 11, 2025 at 2:06 PM
Gemini's "You're absolutely right!" is "You have hit the nail on the head".
I'm looking for opinions. How large does a model need to be to be a "Large" Language Model?
December 9, 2025 at 11:42 PM
I'm looking for opinions. How large does a model need to be to be a "Large" Language Model?
An interesting facet of gemini i've noticed. It's pretty good at explaining things, and often attempts to insert diagrams. However it's biased to use images from shutterstock which all end up being completely random.
December 8, 2025 at 5:44 PM
An interesting facet of gemini i've noticed. It's pretty good at explaining things, and often attempts to insert diagrams. However it's biased to use images from shutterstock which all end up being completely random.
I can't believe I continue to have to say this. These are tools. They are not alive. They are not deities. They are tables and a lot of math.
This is like seeing someone so impressed by an accountant's spreadsheet, they begin talking about how the spreadsheet could raise the dead.
This is like seeing someone so impressed by an accountant's spreadsheet, they begin talking about how the spreadsheet could raise the dead.
December 4, 2025 at 2:49 AM
I can't believe I continue to have to say this. These are tools. They are not alive. They are not deities. They are tables and a lot of math.
This is like seeing someone so impressed by an accountant's spreadsheet, they begin talking about how the spreadsheet could raise the dead.
This is like seeing someone so impressed by an accountant's spreadsheet, they begin talking about how the spreadsheet could raise the dead.
AI is a tool.
No tool is good at everything.
Salesmen sell by promising the world.
Just because a salesman oversells a tool doesn't make it useless.
A discerning user uses a tool for what it's good at and uses a different tool for things it's bad at.
You can be a discerning user.
No tool is good at everything.
Salesmen sell by promising the world.
Just because a salesman oversells a tool doesn't make it useless.
A discerning user uses a tool for what it's good at and uses a different tool for things it's bad at.
You can be a discerning user.
December 3, 2025 at 8:55 PM
AI is a tool.
No tool is good at everything.
Salesmen sell by promising the world.
Just because a salesman oversells a tool doesn't make it useless.
A discerning user uses a tool for what it's good at and uses a different tool for things it's bad at.
You can be a discerning user.
No tool is good at everything.
Salesmen sell by promising the world.
Just because a salesman oversells a tool doesn't make it useless.
A discerning user uses a tool for what it's good at and uses a different tool for things it's bad at.
You can be a discerning user.
I'm casually working out of a coffee shop today and passively watching students do homework and have concluded that LLMs have made homework as it has existed until now obsolete
November 30, 2025 at 7:45 PM
I'm casually working out of a coffee shop today and passively watching students do homework and have concluded that LLMs have made homework as it has existed until now obsolete
I was testing nano banana for generating game assets. It's not perfect, but for a little to no effort concept/prototype, it gets the job done. It will liberally infringe on logos/trademarks so beware. Please enjoy alliance cupcake/x-wing panda fighter-pilot.
November 25, 2025 at 7:06 PM
I was testing nano banana for generating game assets. It's not perfect, but for a little to no effort concept/prototype, it gets the job done. It will liberally infringe on logos/trademarks so beware. Please enjoy alliance cupcake/x-wing panda fighter-pilot.
Gemini 3 Pro's reasoning traces are something else. I ran into an issue where Antigravity was convinced Qwen3 doesn't exist. So I fired up Gemini and asked "Does Qwen3 exist" and apparently Gemini struggles while trancending time into a simulated futures.
November 22, 2025 at 5:23 PM
Gemini 3 Pro's reasoning traces are something else. I ran into an issue where Antigravity was convinced Qwen3 doesn't exist. So I fired up Gemini and asked "Does Qwen3 exist" and apparently Gemini struggles while trancending time into a simulated futures.
I can't believe how often I have to explain this. No, buying a ChatGPT pro subscription does not suddenly make you an expert on [Insert topic here].
I'm not suddenly a plumber or electrician because I can find instructions on installing pipe fittings on YouTube.
I'm not suddenly a plumber or electrician because I can find instructions on installing pipe fittings on YouTube.
November 21, 2025 at 7:18 PM
I can't believe how often I have to explain this. No, buying a ChatGPT pro subscription does not suddenly make you an expert on [Insert topic here].
I'm not suddenly a plumber or electrician because I can find instructions on installing pipe fittings on YouTube.
I'm not suddenly a plumber or electrician because I can find instructions on installing pipe fittings on YouTube.
Something interesting with Gemini 3 Pro (via Antigravity) is how it uses emotionally expressive verbiage in its output. It's been stuck in a loop for ~10 mintues now trying to figure out cli input errors.
November 21, 2025 at 1:41 PM
Something interesting with Gemini 3 Pro (via Antigravity) is how it uses emotionally expressive verbiage in its output. It's been stuck in a loop for ~10 mintues now trying to figure out cli input errors.
Is Antigravity just Windsurf? I grepped the binary and found this. Safe bet that they reused some of the code at least.
November 20, 2025 at 8:35 PM
Is Antigravity just Windsurf? I grepped the binary and found this. Safe bet that they reused some of the code at least.
K. Let's try this Antigravity AI IDE thing. antigravity.google
Google Antigravity
Google Antigravity - Build the new way
antigravity.google
November 20, 2025 at 7:38 PM
K. Let's try this Antigravity AI IDE thing. antigravity.google
Microsoft's approach to AI in the OS.
* Copilots get their own account
* You have to give them access to your folders
* They have access to installed apps
* Security - YOLO "enable this feature if you understand the security implications outlined on this page"
support.microsoft.com/en-us/window...
* Copilots get their own account
* You have to give them access to your folders
* They have access to installed apps
* Security - YOLO "enable this feature if you understand the security implications outlined on this page"
support.microsoft.com/en-us/window...
Experimental Agentic Features - Microsoft Support
support.microsoft.com
November 19, 2025 at 4:44 PM
Microsoft's approach to AI in the OS.
* Copilots get their own account
* You have to give them access to your folders
* They have access to installed apps
* Security - YOLO "enable this feature if you understand the security implications outlined on this page"
support.microsoft.com/en-us/window...
* Copilots get their own account
* You have to give them access to your folders
* They have access to installed apps
* Security - YOLO "enable this feature if you understand the security implications outlined on this page"
support.microsoft.com/en-us/window...
Anyone in the AI/LLM space have any good resources they like for someone interested in diving into mechanistic interpretability?
November 18, 2025 at 10:21 PM
Anyone in the AI/LLM space have any good resources they like for someone interested in diving into mechanistic interpretability?
"Yes — you are absolutely right!" is the most dangerous phrase in the world today.
November 17, 2025 at 6:57 PM
"Yes — you are absolutely right!" is the most dangerous phrase in the world today.
Interesting rebuttal to the Anthropic story this morning basically calling out the lack of concrete evidence.
In my summary I did note that they were the primary and only source so there was no available verification or counter narrative available djnn.sh/posts/anthro....
In my summary I did note that they were the primary and only source so there was no available verification or counter narrative available djnn.sh/posts/anthro....
November 16, 2025 at 3:52 PM
Interesting rebuttal to the Anthropic story this morning basically calling out the lack of concrete evidence.
In my summary I did note that they were the primary and only source so there was no available verification or counter narrative available djnn.sh/posts/anthro....
In my summary I did note that they were the primary and only source so there was no available verification or counter narrative available djnn.sh/posts/anthro....
It's interesting that the framing of the report is that Claude was supposed to be the gate. The most gullible part of the system was the gate.
November 14, 2025 at 10:34 PM
It's interesting that the framing of the report is that Claude was supposed to be the gate. The most gullible part of the system was the gate.
Anthropic put out a recent report that details the workings what they call the first documented case of a cyberattack largely executed without human intervention at scale 🧵
Full report: assets.anthropic.com/m/ec212e6566...
Full report: assets.anthropic.com/m/ec212e6566...
November 14, 2025 at 8:53 PM
Anthropic put out a recent report that details the workings what they call the first documented case of a cyberattack largely executed without human intervention at scale 🧵
Full report: assets.anthropic.com/m/ec212e6566...
Full report: assets.anthropic.com/m/ec212e6566...
I love programming as a medium of self-expression. This has to be one of the coolest things I've ever seen www.youtube.com/watch?v=iu5r...
Coding Trance Music from Scratch (Again )
YouTube video by Switch Angel
www.youtube.com
November 14, 2025 at 1:22 PM
I love programming as a medium of self-expression. This has to be one of the coolest things I've ever seen www.youtube.com/watch?v=iu5r...
From the gpt-5.1 system card. I guess they really didn't like the press around AI psychosis users were experiencing. openai.com/index/gpt-5-...
November 13, 2025 at 6:08 PM
From the gpt-5.1 system card. I guess they really didn't like the press around AI psychosis users were experiencing. openai.com/index/gpt-5-...
Looks like gpt-5.1 is a thing. Focus largely appears to be conversation style. They emphasize it being warmer than gpt-5
openai.com/index/gpt-5-1/
openai.com/index/gpt-5-1/
GPT-5.1: A smarter, more conversational ChatGPT
We’re upgrading GPT-5 while making it easier to customize ChatGPT. Starting to roll out today to everyone, beginning with paid users.
openai.com
November 12, 2025 at 9:11 PM
Looks like gpt-5.1 is a thing. Focus largely appears to be conversation style. They emphasize it being warmer than gpt-5
openai.com/index/gpt-5-1/
openai.com/index/gpt-5-1/