Nish Tahir
@nishtahir.com
Principal Engineer, ML/AI Research (Anti-hype). I can and will be wrong. My opinions are my own.
Blog: https://nishtahir.com
Mastodon: social.nishtahir.com/@nish
Blog: https://nishtahir.com
Mastodon: social.nishtahir.com/@nish
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/
It's unfortunate that all I keep hearing from Windows users is just how bad ads in the OS have gotten. My only use case is Steam big picture because Halo kept panicking on Bazzite. I don't even see the desktop.
arstechnica.com/gadgets/2025...
arstechnica.com/gadgets/2025...
How to declutter, quiet down, and take the AI out of Windows 11 25H2
A new major Windows 11 release means a new guide for cleaning up the OS.
arstechnica.com
November 6, 2025 at 10:22 PM
It's unfortunate that all I keep hearing from Windows users is just how bad ads in the OS have gotten. My only use case is Steam big picture because Halo kept panicking on Bazzite. I don't even see the desktop.
arstechnica.com/gadgets/2025...
arstechnica.com/gadgets/2025...
I didn't have space datacenters on my 2025 AI hypecycle bingo arstechnica.com/google/2025/...
Meet Project Suncatcher, Google’s plan to put AI data centers in space
Google is already zapping TPUs with radiation to get ready.
arstechnica.com
November 4, 2025 at 10:37 PM
I didn't have space datacenters on my 2025 AI hypecycle bingo arstechnica.com/google/2025/...
me talking to some of my younger engineers about IRC and everyone responds with "tf is that?"
a close up of a man in a military uniform looking at the camera with a serious look on his face .
Alt: a close up of a man in a military uniform looking at the camera with a serious look on his face aging rapidly into an old man.
media.tenor.com
November 4, 2025 at 10:07 PM
me talking to some of my younger engineers about IRC and everyone responds with "tf is that?"
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...
We really missed an opportunity to make all the LLM acronyms valid roman numerals. LLM, VLLM, LVM, LMMLM...
October 28, 2025 at 5:18 PM
We really missed an opportunity to make all the LLM acronyms valid roman numerals. LLM, VLLM, LVM, LMMLM...
Looks like they are delegating more responsibility to the tool. Hopefully watch dog groups keep an eye on this to identify issues with hallucinations and bias.
indianexpress.com/article/worl...
indianexpress.com/article/worl...
Albania’s AI minister ‘pregnant’ with 83 children, says PM Edi Rama
First launched in January as a virtual assistant on the e-Albania platform, Diella helps citizens and businesses access state documents.
indianexpress.com
October 27, 2025 at 1:07 PM
Looks like they are delegating more responsibility to the tool. Hopefully watch dog groups keep an eye on this to identify issues with hallucinations and bias.
indianexpress.com/article/worl...
indianexpress.com/article/worl...
I got an ad from Virgin Voyages that looked like something straight out of stable diffusion. Turns out they've been running AI generated ad campaigns. A lot of these seem very generic so I wonder how well they work for them.
adage.com/brand-market...
adage.com/brand-market...
Inside Virgin Voyages’ agentic AI strategy that’s slashing its agency costs
The brand is using more than 50 agents to personalize marketing and help guide consumers toward booking.
adage.com
October 25, 2025 at 10:56 PM
I got an ad from Virgin Voyages that looked like something straight out of stable diffusion. Turns out they've been running AI generated ad campaigns. A lot of these seem very generic so I wonder how well they work for them.
adage.com/brand-market...
adage.com/brand-market...
Everyone should be familiar with the DNS haiku
October 23, 2025 at 9:42 PM
Everyone should be familiar with the DNS haiku
This is interesting. Google released a guided prompt eval/optimization tool unsurprisingly based around vertex. What I was not expecting was for it to be using streamlit. github.com/GoogleCloudP...
generative-ai/tools/llmevalkit at main · GoogleCloudPlatform/generative-ai
Sample code and notebooks for Generative AI on Google Cloud, with Gemini on Vertex AI - GoogleCloudPlatform/generative-ai
github.com
October 23, 2025 at 8:05 PM
This is interesting. Google released a guided prompt eval/optimization tool unsurprisingly based around vertex. What I was not expecting was for it to be using streamlit. github.com/GoogleCloudP...
Another line in the bubble chart I guess? finance.yahoo.com/news/anthrop...
Anthropic, Google in Talks on Multibillion-Dollar Cloud Deal
The plan, which has not been finalized, involves Google providing cloud computing services to Anthropic, according to the people, who asked not to be named because the information is private. The dea...
finance.yahoo.com
October 23, 2025 at 6:42 PM
Another line in the bubble chart I guess? finance.yahoo.com/news/anthrop...
And OpenAI released a browser. The race to mine browsing habbits for context is on.
techcrunch.com/2025/10/21/o...
techcrunch.com/2025/10/21/o...
OpenAI launches an AI-powered browser: ChatGPT Atlas | TechCrunch
OpenAI is launching an AI-powered browser, its latest challenge to Google as the main way people find information online.
techcrunch.com
October 21, 2025 at 5:36 PM
And OpenAI released a browser. The race to mine browsing habbits for context is on.
techcrunch.com/2025/10/21/o...
techcrunch.com/2025/10/21/o...
I wrote about the DORA State of AI Assisted software engineering report nishtahir.com/on-the-dora-...
On the DORA 2025 AI Report - AI Adoption and use
The DORA State of AI Assisted Software Development report came out recently. It's a massive 142 page report that details analysis done by the DORA team. It captures trends and observations covering a ...
nishtahir.com
October 21, 2025 at 3:16 AM
I wrote about the DORA State of AI Assisted software engineering report nishtahir.com/on-the-dora-...
Speculation in AI has always been and continues to be a waste of time. Self driving was 2 years away 10 years ago. AGI was 2 years away 2 years ago and is now 10 years away.
Focus on what actually works right now and build against that. There's a lot of cool stuff already here to explore.
Focus on what actually works right now and build against that. There's a lot of cool stuff already here to explore.
October 20, 2025 at 1:25 AM
Speculation in AI has always been and continues to be a waste of time. Self driving was 2 years away 10 years ago. AGI was 2 years away 2 years ago and is now 10 years away.
Focus on what actually works right now and build against that. There's a lot of cool stuff already here to explore.
Focus on what actually works right now and build against that. There's a lot of cool stuff already here to explore.
IMO unless you specifically care about CUDA, getting AMD's AI Max 395+ is a much better deal. The currently show better performance and you can get 2 nodes for the price of 1 DGX spark.
www.inc.com/chris-morris...
www.inc.com/chris-morris...
Why Your Business Should Consider Nvidia's New $4,000 Desktop Supercomputer
The DGX Spark supercomputer just went on sale. Here's what to know about the powerful machine.
www.inc.com
October 18, 2025 at 9:32 PM
IMO unless you specifically care about CUDA, getting AMD's AI Max 395+ is a much better deal. The currently show better performance and you can get 2 nodes for the price of 1 DGX spark.
www.inc.com/chris-morris...
www.inc.com/chris-morris...
Claude Skills appear to be little pre written bits of context that Claude can include when attempting to do stuff. There's currently a comparison to MCP but these are 2 different things
www.anthropic.com/news/skills
www.anthropic.com/news/skills
Claude Skills: Customize AI for your workflows
Build custom Skills to teach Claude specialized tasks. Create once, use everywhere—from spreadsheets to coding. Available across Claude.ai, API, and Code.
www.anthropic.com
October 17, 2025 at 10:24 PM
Claude Skills appear to be little pre written bits of context that Claude can include when attempting to do stuff. There's currently a comparison to MCP but these are 2 different things
www.anthropic.com/news/skills
www.anthropic.com/news/skills
The competing products in the space for Large model local inference are the Mac Mini/Pro line and the AMD AI Max line
llm-tracker.info/AMD-Strix-Ha...
llm-tracker.info/AMD-Strix-Ha...
AMD Strix Halo (Ryzen AI Max+ 395) GPU Performance
I’ve been doing some (ongoing) testing on a Strix Halo system and with a bunch of desktop systems coming out soon, I figured it might be worth sharing a few notes of on the current performance.
llm-tracker.info
October 15, 2025 at 3:52 PM
The competing products in the space for Large model local inference are the Mac Mini/Pro line and the AMD AI Max line
llm-tracker.info/AMD-Strix-Ha...
llm-tracker.info/AMD-Strix-Ha...
Super exciting times ahead. Reviews of the DGX Spark are starting to turn up. Overall looks like high high-bandwidth unified CPU/GPU memory architecture that suffers from low compute performance.
Good for edge/prototyping but not prod workloads
lmsys.org/blog/2025-10...
Good for edge/prototyping but not prod workloads
lmsys.org/blog/2025-10...
NVIDIA DGX Spark In-Depth Review: A New Standard for Local AI Inference | LMSYS Org
<p>Thanks to NVIDIA’s early access program, we are thrilled to get our hands on the NVIDIA DGX™ Spark. It’s quite an unconventional system, as NVIDIA rarely ...
lmsys.org
October 15, 2025 at 3:31 PM
Super exciting times ahead. Reviews of the DGX Spark are starting to turn up. Overall looks like high high-bandwidth unified CPU/GPU memory architecture that suffers from low compute performance.
Good for edge/prototyping but not prod workloads
lmsys.org/blog/2025-10...
Good for edge/prototyping but not prod workloads
lmsys.org/blog/2025-10...
I'm convinced that Phoenix has to be the most common but perfect project name of all time.
You don't even have to ask any questions. Once you get put on "Phoenix" you know exactly what happened.
You don't even have to ask any questions. Once you get put on "Phoenix" you know exactly what happened.
October 14, 2025 at 10:53 PM
I'm convinced that Phoenix has to be the most common but perfect project name of all time.
You don't even have to ask any questions. Once you get put on "Phoenix" you know exactly what happened.
You don't even have to ask any questions. Once you get put on "Phoenix" you know exactly what happened.
Apple TV+ rebranding to Apple TV. Not to be confused with Apple TV - the app, or Apple TV - the set top box. Yo dawg, I heard you like Apple TV...
www.theverge.com/news/798857/...
www.theverge.com/news/798857/...
Apple TV Plus is being rebranded to… Apple TV
Apple says it’s given the streaming service a ‘vibrant new identity.’
www.theverge.com
October 13, 2025 at 7:46 PM
Apple TV+ rebranding to Apple TV. Not to be confused with Apple TV - the app, or Apple TV - the set top box. Yo dawg, I heard you like Apple TV...
www.theverge.com/news/798857/...
www.theverge.com/news/798857/...
LLMs are not inherently trustworthy. As they get integrated into tools workflows, information sources a good stance is to assume they are untrustworthy for any given job until they prove otherwise
www.anthropic.com/research/sma...
www.anthropic.com/research/sma...
A small number of samples can poison LLMs of any size
Anthropic research on data-poisoning attacks in large language models
www.anthropic.com
October 10, 2025 at 1:13 AM
LLMs are not inherently trustworthy. As they get integrated into tools workflows, information sources a good stance is to assume they are untrustworthy for any given job until they prove otherwise
www.anthropic.com/research/sma...
www.anthropic.com/research/sma...
OpenAI dev day summary
* 2025 usage stats show they had 4M devs and burn 6B tokens/min
* ChatGPT apps. You can now talk to Zillow or Canva and ask them to do stuff - openai.com/index/introd...
* If you expose an MCP server, you can expose your agent to ChatGPT and target that audience
1/2
* 2025 usage stats show they had 4M devs and burn 6B tokens/min
* ChatGPT apps. You can now talk to Zillow or Canva and ask them to do stuff - openai.com/index/introd...
* If you expose an MCP server, you can expose your agent to ChatGPT and target that audience
1/2
Introducing apps in ChatGPT and the new Apps SDK
A new generation of apps you can chat with and the tools for developers to build them.
openai.com
October 7, 2025 at 2:04 AM
OpenAI dev day summary
* 2025 usage stats show they had 4M devs and burn 6B tokens/min
* ChatGPT apps. You can now talk to Zillow or Canva and ask them to do stuff - openai.com/index/introd...
* If you expose an MCP server, you can expose your agent to ChatGPT and target that audience
1/2
* 2025 usage stats show they had 4M devs and burn 6B tokens/min
* ChatGPT apps. You can now talk to Zillow or Canva and ask them to do stuff - openai.com/index/introd...
* If you expose an MCP server, you can expose your agent to ChatGPT and target that audience
1/2