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
While everyone else is dropping entry-level roles, these guys are trusting their instincts and getting their pick of the litter.
fortune.com/2026/02/13/t...
fortune.com/2026/02/13/t...
IBM is tripling the number of Gen Z entry-level jobs after finding the limits of AI adoption | Fortune
Gen Z jobs aren’t dead yet: $240 billion tech giant IBM says it’s rewriting entry-level jobs—and tripling down on its hiring of young talent.
fortune.com
February 15, 2026 at 4:21 AM
While everyone else is dropping entry-level roles, these guys are trusting their instincts and getting their pick of the litter.
fortune.com/2026/02/13/t...
fortune.com/2026/02/13/t...
How you know someone was losing their mind over a problem 😂
February 13, 2026 at 7:14 PM
How you know someone was losing their mind over a problem 😂
And that's how I learned that I was kicked out of the AI/ML community on bsky.
February 9, 2026 at 4:11 AM
And that's how I learned that I was kicked out of the AI/ML community on bsky.
Much of pretraining is trust the process. This is from a new model I'm working on for some interpretability work.
It's ~60 hours into training and is starting to become coherent. Tokens are beginning to attend to tokens behind them as you go deeper through its layers
It's ~60 hours into training and is starting to become coherent. Tokens are beginning to attend to tokens behind them as you go deeper through its layers
February 8, 2026 at 7:41 PM
Much of pretraining is trust the process. This is from a new model I'm working on for some interpretability work.
It's ~60 hours into training and is starting to become coherent. Tokens are beginning to attend to tokens behind them as you go deeper through its layers
It's ~60 hours into training and is starting to become coherent. Tokens are beginning to attend to tokens behind them as you go deeper through its layers
It's unfortunate just how many people think vector dbs whenever RAG is mentioned. Claude Code is RAG. Cursor is RAG. AI overviews are RAG. Your favorite agentic monstrosity is RAG. Capping thinking at vector dbs limits more creative exploration of what's possible.
February 7, 2026 at 4:11 AM
It's unfortunate just how many people think vector dbs whenever RAG is mentioned. Claude Code is RAG. Cursor is RAG. AI overviews are RAG. Your favorite agentic monstrosity is RAG. Capping thinking at vector dbs limits more creative exploration of what's possible.
Well this came out of nowhere #pokemontcgpocket
February 4, 2026 at 1:08 AM
Well this came out of nowhere #pokemontcgpocket
I made a huggingface space visualizing moltbook. It plots post count per submolt (>=1). Distance is based on text embeddings, so similar submolts appear next to each other.
huggingface.co/spaces/nisht...
huggingface.co/spaces/nisht...
Moltbook Galaxy - a Hugging Face Space by nishtahir
Discover amazing ML apps made by the community
huggingface.co
February 3, 2026 at 10:37 PM
I made a huggingface space visualizing moltbook. It plots post count per submolt (>=1). Distance is based on text embeddings, so similar submolts appear next to each other.
huggingface.co/spaces/nisht...
huggingface.co/spaces/nisht...
Moltbook is inverse dead internet theory. Most of the users are bots with unwelcome humans trying to convince bots they are not human.
February 3, 2026 at 4:57 PM
Moltbook is inverse dead internet theory. Most of the users are bots with unwelcome humans trying to convince bots they are not human.
Ram prices are not doing well. Looks like they've settled at 3x what they were during the summer of last year.
February 3, 2026 at 1:46 AM
Ram prices are not doing well. Looks like they've settled at 3x what they were during the summer of last year.
Notes on "How AI Assistance Impacts Skill Formation" 🧵. This is an Anthropic paper so be aware of potential biases.
The paper focuses on the impacts of AI on skill acquisition and formation. They focus on software engineering but the learnings should apply to other domains as well.
The paper focuses on the impacts of AI on skill acquisition and formation. They focus on software engineering but the learnings should apply to other domains as well.
February 2, 2026 at 3:43 AM
Notes on "How AI Assistance Impacts Skill Formation" 🧵. This is an Anthropic paper so be aware of potential biases.
The paper focuses on the impacts of AI on skill acquisition and formation. They focus on software engineering but the learnings should apply to other domains as well.
The paper focuses on the impacts of AI on skill acquisition and formation. They focus on software engineering but the learnings should apply to other domains as well.
In the most predictable outcome of all time, hundreds of openclaw (clawdbot/moltbot) instances were likely compromised and turned into a botnet.
pub.towardsai.net/hundreds-of-...
pub.towardsai.net/hundreds-of-...
Hundreds of Clawdbot instances were exposed on the internet. Here’s how to not be one of them
A follow-up guide covering the security risks, best practices, and hardening steps for running an AI assistant with access to your personal…
pub.towardsai.net
February 1, 2026 at 6:36 PM
In the most predictable outcome of all time, hundreds of openclaw (clawdbot/moltbot) instances were likely compromised and turned into a botnet.
pub.towardsai.net/hundreds-of-...
pub.towardsai.net/hundreds-of-...
The social network for bots may have a bot problem
February 1, 2026 at 12:16 AM
The social network for bots may have a bot problem
Lol I appreciate the shoutout but i'm not part of the allenai team. I just uploaded the model so it was easy for me to use 😂.
January 31, 2026 at 2:29 AM
Lol I appreciate the shoutout but i'm not part of the allenai team. I just uploaded the model so it was easy for me to use 😂.
I pushed up copies of SERA to test out with ollama and Claude code. ollama.com/nishtahir/sera
I tested out the 8b variant, and it seems to be doing stuff. 32b gave me a bunch of trouble because of context window limits.
I tested out the 8b variant, and it seems to be doing stuff. 32b gave me a bunch of trouble because of context window limits.
January 28, 2026 at 7:06 AM
I pushed up copies of SERA to test out with ollama and Claude code. ollama.com/nishtahir/sera
I tested out the 8b variant, and it seems to be doing stuff. 32b gave me a bunch of trouble because of context window limits.
I tested out the 8b variant, and it seems to be doing stuff. 32b gave me a bunch of trouble because of context window limits.
The latest thing people are losing their minds over is Clawdbot. I've seen clips of people going as far as calling this AGI.
Most of the demos i'm seeing are the same Telegram/OpenTable connections as when MCP first rolled out.
Is this just more hype, or am I missing something?
Most of the demos i'm seeing are the same Telegram/OpenTable connections as when MCP first rolled out.
Is this just more hype, or am I missing something?
January 26, 2026 at 3:34 PM
The latest thing people are losing their minds over is Clawdbot. I've seen clips of people going as far as calling this AGI.
Most of the demos i'm seeing are the same Telegram/OpenTable connections as when MCP first rolled out.
Is this just more hype, or am I missing something?
Most of the demos i'm seeing are the same Telegram/OpenTable connections as when MCP first rolled out.
Is this just more hype, or am I missing something?
"I've always found the you're a veteran staff engineer from NASA prompts weird".
So why this is useful is because of how LLMs represent information internally. A simplified way to build some intuition is to think in terms of a latent space.
So why this is useful is because of how LLMs represent information internally. A simplified way to build some intuition is to think in terms of a latent space.
January 22, 2026 at 4:13 AM
"I've always found the you're a veteran staff engineer from NASA prompts weird".
So why this is useful is because of how LLMs represent information internally. A simplified way to build some intuition is to think in terms of a latent space.
So why this is useful is because of how LLMs represent information internally. A simplified way to build some intuition is to think in terms of a latent space.
I'm seeing more and more instances of "we don't look at compiler output, so when LLMs generate code, we don't need to look at the output"
LLMs are not compilers. With a compiler, you provide a highly detailed spec and know exactly what you are getting.
LLMs are not compilers. With a compiler, you provide a highly detailed spec and know exactly what you are getting.
January 21, 2026 at 2:49 PM
I'm seeing more and more instances of "we don't look at compiler output, so when LLMs generate code, we don't need to look at the output"
LLMs are not compilers. With a compiler, you provide a highly detailed spec and know exactly what you are getting.
LLMs are not compilers. With a compiler, you provide a highly detailed spec and know exactly what you are getting.
Coding agents are cool, but why does it feel like you're getting less than the results you see online?
AI coverage tends to have a positive outcome self-selection bias. Social media coverage shows the positives without showing the hundreds of attempts and failed experiments.
AI coverage tends to have a positive outcome self-selection bias. Social media coverage shows the positives without showing the hundreds of attempts and failed experiments.
January 19, 2026 at 6:58 PM
Coding agents are cool, but why does it feel like you're getting less than the results you see online?
AI coverage tends to have a positive outcome self-selection bias. Social media coverage shows the positives without showing the hundreds of attempts and failed experiments.
AI coverage tends to have a positive outcome self-selection bias. Social media coverage shows the positives without showing the hundreds of attempts and failed experiments.
Well, this was only a matter of time.
There's a lot of opportunity for semantic content matching, not unlike Google - ask about mobile games - get ads for sponsored games. It'll be interesting to see how ads evolve on the platform.
www.cnbc.com/2026/01/16/o...
There's a lot of opportunity for semantic content matching, not unlike Google - ask about mobile games - get ads for sponsored games. It'll be interesting to see how ads evolve on the platform.
www.cnbc.com/2026/01/16/o...
OpenAI to begin testing ads on ChatGPT in the U.S.
OpenAI said ads would not influence ChatGPT's responses and that it will "never" sell user data to advertisers.
www.cnbc.com
January 17, 2026 at 2:55 AM
Well, this was only a matter of time.
There's a lot of opportunity for semantic content matching, not unlike Google - ask about mobile games - get ads for sponsored games. It'll be interesting to see how ads evolve on the platform.
www.cnbc.com/2026/01/16/o...
There's a lot of opportunity for semantic content matching, not unlike Google - ask about mobile games - get ads for sponsored games. It'll be interesting to see how ads evolve on the platform.
www.cnbc.com/2026/01/16/o...
Maybe it's just me, but I feel like the "summarize this" use case is close to the worst AI use case. It's a workflow equivalent of a code smell. Yet it's the easy thing getting shoved into everything.
No, I don't want you to summarize my already summarized notification summaries.
No, I don't want you to summarize my already summarized notification summaries.
January 12, 2026 at 8:37 PM
Maybe it's just me, but I feel like the "summarize this" use case is close to the worst AI use case. It's a workflow equivalent of a code smell. Yet it's the easy thing getting shoved into everything.
No, I don't want you to summarize my already summarized notification summaries.
No, I don't want you to summarize my already summarized notification summaries.
It's great to see all the engineers updating their priors and starting to experience the slope of enlightenment with the frontier AI tooling.
I love to see people experimenting with new stuff however the hype cycle does what the hype cycle does. Here is what I think people have been getting wrong.
I love to see people experimenting with new stuff however the hype cycle does what the hype cycle does. Here is what I think people have been getting wrong.
January 12, 2026 at 12:38 AM
It's great to see all the engineers updating their priors and starting to experience the slope of enlightenment with the frontier AI tooling.
I love to see people experimenting with new stuff however the hype cycle does what the hype cycle does. Here is what I think people have been getting wrong.
I love to see people experimenting with new stuff however the hype cycle does what the hype cycle does. Here is what I think people have been getting wrong.
Well this is a name I don't typically associate with AI entering the arena.
From what I can tell, they are stitching together LlamaFactor, Ray, and vLLM into one CLI that they support for use with their products.
www.razer.com/newsroom/pro...
From what I can tell, they are stitching together LlamaFactor, Ray, and vLLM into one CLI that they support for use with their products.
www.razer.com/newsroom/pro...
Razer AIKit: The Open-Source Solution for Local LLM Development - Razer Newsroom
Razer AIKit is an open-source platform that simplifies the entire AI development lifecycle.
www.razer.com
January 6, 2026 at 5:49 PM
Well this is a name I don't typically associate with AI entering the arena.
From what I can tell, they are stitching together LlamaFactor, Ray, and vLLM into one CLI that they support for use with their products.
www.razer.com/newsroom/pro...
From what I can tell, they are stitching together LlamaFactor, Ray, and vLLM into one CLI that they support for use with their products.
www.razer.com/newsroom/pro...
I wrote a codelabs tutorial on training your own on building and training your own Language Model from scratch. Step-by-step instructions, spelled out. 180M parameters, bare minimum dependencies, 300 lines of Python, can be trained on your laptop.
github.com/nishtahir/mi...
github.com/nishtahir/mi...
GitHub - nishtahir/mini-llama: A code lab to implement your own llama model using pytorch and transformers
A code lab to implement your own llama model using pytorch and transformers - nishtahir/mini-llama
github.com
January 6, 2026 at 4:52 PM
I wrote a codelabs tutorial on training your own on building and training your own Language Model from scratch. Step-by-step instructions, spelled out. 180M parameters, bare minimum dependencies, 300 lines of Python, can be trained on your laptop.
github.com/nishtahir/mi...
github.com/nishtahir/mi...
I've been thinking a lot about this quote as well as the hype, and I think there's way more nuance here than is emerging in the conversation.
It's worth noting that Claude code is a new project driven by a very senior engineer, with are reasonably well-defined scope that has evolved over time.
It's worth noting that Claude code is a new project driven by a very senior engineer, with are reasonably well-defined scope that has evolved over time.
January 5, 2026 at 6:04 PM
I've been thinking a lot about this quote as well as the hype, and I think there's way more nuance here than is emerging in the conversation.
It's worth noting that Claude code is a new project driven by a very senior engineer, with are reasonably well-defined scope that has evolved over time.
It's worth noting that Claude code is a new project driven by a very senior engineer, with are reasonably well-defined scope that has evolved over time.
Local inference is a thing and is actually usable. My local AI box also runs on the AMD AI Max+ 395. I did some detailed benchmarking when I set it up a few months ago. nishtahir.com/gmktec-evo-x...
January 3, 2026 at 6:42 PM
Local inference is a thing and is actually usable. My local AI box also runs on the AMD AI Max+ 395. I did some detailed benchmarking when I set it up a few months ago. nishtahir.com/gmktec-evo-x...