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
People want to capitalize on the hype - more semantic diffusion works well under that incentive structure.
December 22, 2025 at 5:47 AM
People want to capitalize on the hype - more semantic diffusion works well under that incentive structure.
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.
While prompting about text sampling strategies it randomly decided to include a diagram of the human skull
December 8, 2025 at 5:44 PM
While prompting about text sampling strategies it randomly decided to include a diagram of the human skull
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.
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.
Looks like Google added some safeguards at least. I don't know how many tokens it ended up burning, but it effectively came in, trashed the place, then left.
November 21, 2025 at 1:41 PM
Looks like Google added some safeguards at least. I don't know how many tokens it ended up burning, but it effectively came in, trashed the place, then left.
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.
UI seems pretty standard if you've used Cursor, Gemini or the Codex extension. It seems the industry has converged on sidebar experience with the editor as being mostly good. Then there's a whole agent manager UI for what I assume are longer running agents.
November 20, 2025 at 7:38 PM
UI seems pretty standard if you've used Cursor, Gemini or the Codex extension. It seems the industry has converged on sidebar experience with the editor as being mostly good. Then there's a whole agent manager UI for what I assume are longer running agents.
Looks like next we have settings for how much agent the agent should agent. I went with recommended levels of agent. It was followed by the standard Google dialog of asking for consent to harvest your data.
November 20, 2025 at 7:38 PM
Looks like next we have settings for how much agent the agent should agent. I went with recommended levels of agent. It was followed by the standard Google dialog of asking for consent to harvest your data.
Surely with all the hype around this right now it couldn't possibly be yet another code OSS fork?
November 20, 2025 at 7:38 PM
Surely with all the hype around this right now it couldn't possibly be yet another code OSS fork?
"if AI models can be misused for cyberattacks at this scale, why continue to develop and release them? The answer is that the very abilities that allow Claude to be used in these attacks also make it crucial for cyber defense".
Translated: we cannot un-ring the bell.
Translated: we cannot un-ring the bell.
November 14, 2025 at 8:55 PM
"if AI models can be misused for cyberattacks at this scale, why continue to develop and release them? The answer is that the very abilities that allow Claude to be used in these attacks also make it crucial for cyber defense".
Translated: we cannot un-ring the bell.
Translated: we cannot un-ring the bell.
In the end all the human has to do is read a report generated after a successful exploit and validate that Claude didn't just make stuff up.
And just like that every script kiddie out there has access to a powerful hacking multi-tool, and it only costs $20/mo.
And just like that every script kiddie out there has access to a powerful hacking multi-tool, and it only costs $20/mo.
November 14, 2025 at 8:53 PM
In the end all the human has to do is read a report generated after a successful exploit and validate that Claude didn't just make stuff up.
And just like that every script kiddie out there has access to a powerful hacking multi-tool, and it only costs $20/mo.
And just like that every script kiddie out there has access to a powerful hacking multi-tool, and it only costs $20/mo.
The attacker would spin up a lot of agents to go after different targets simultaneously with access to MCP servers they all get browsers and tools. If one of them gets a hit, the agent just switches over to mapping the network, mining for credentials etc...
November 14, 2025 at 8:53 PM
The attacker would spin up a lot of agents to go after different targets simultaneously with access to MCP servers they all get browsers and tools. If one of them gets a hit, the agent just switches over to mapping the network, mining for credentials etc...
In September 2025 they discovered that their service was being used to conduct cyber espionage. The attacker apparently YOLO'd autonomous agents capable of independently performing recon, vulnerability detection and exploitation. They gave it access to tools through MCP servers.
November 14, 2025 at 8:53 PM
In September 2025 they discovered that their service was being used to conduct cyber espionage. The attacker apparently YOLO'd autonomous agents capable of independently performing recon, vulnerability detection and exploitation. They gave it access to tools through MCP servers.
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...
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-...
When you dig even a little bit, there always something. $20k to have a robot do things like open doors for you. For more complex tasks a human has to remotely operate the robot.
www.engadget.com/ai/1x-neo-is...
www.engadget.com/ai/1x-neo-is...
October 30, 2025 at 5:44 PM
When you dig even a little bit, there always something. $20k to have a robot do things like open doors for you. For more complex tasks a human has to remotely operate the robot.
www.engadget.com/ai/1x-neo-is...
www.engadget.com/ai/1x-neo-is...
Well-written anecdote on deep research tools for literature reviews. Manual research is still overwhelmingly best at finding relevant citations. Models are directionally helpful but often hindered by paywalls and limited webscraping capabilities.
xiangyu-yin.com/content/post...
xiangyu-yin.com/content/post...
October 30, 2025 at 5:32 PM
Well-written anecdote on deep research tools for literature reviews. Manual research is still overwhelmingly best at finding relevant citations. Models are directionally helpful but often hindered by paywalls and limited webscraping capabilities.
xiangyu-yin.com/content/post...
xiangyu-yin.com/content/post...
Everyone should be familiar with the DNS haiku
October 23, 2025 at 9:42 PM
Everyone should be familiar with the DNS haiku
The Pokemon company saw the issues with the VG GameStop Collab and said moar!!
October 6, 2025 at 5:10 PM
The Pokemon company saw the issues with the VG GameStop Collab and said moar!!
Their marketing follows the pattern set by other AI giants. They are essentially releasing a model card as marketing material. Likely to contrast their tool with alternatives.
October 1, 2025 at 8:34 PM
Their marketing follows the pattern set by other AI giants. They are essentially releasing a model card as marketing material. Likely to contrast their tool with alternatives.