Author | Lightnews

Nish Tahir

Nish Tahir

@nishtahir.com

92 followers 46 following 660 posts

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

Posts Media Videos Starter Packs

Nish Tahir @nishtahir.com · 2d

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...

However, if early adopters want NEO to be capable of more specific or complex tasks, they'd have to be comfortable with the idea of a human teleoperator controlling the robot remotely and seeing inside their homes.

Nish Tahir @nishtahir.com · 2d

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...

Nish Tahir @nishtahir.com · 4d

We really missed an opportunity to make all the LLM acronyms valid roman numerals. LLM, VLLM, LVM, LMMLM...

Nish Tahir @nishtahir.com · 5d

It's worth noting that I did some googling and couldn't find anything about the architecture, models, LLM, AI stack. If the goal is transparency I'd have expected this to be publicly available to review.

Nish Tahir @nishtahir.com · 5d

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...

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

Nish Tahir @nishtahir.com · 5d

Welcome new friends. 👋

Nish Tahir @nishtahir.com · 7d

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...

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.

Nish Tahir @nishtahir.com · 9d

Everyone should be familiar with the DNS haiku

Japanese style ink art of a cherry blossom branch with pink flowers. On the top right a haiku reading

It's not DNS
There's no way it's DNS
It was DNS

- SSBroski

Nish Tahir @nishtahir.com · 9d

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

Nish Tahir @nishtahir.com · 9d

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

Nish Tahir @nishtahir.com · 11d

And OpenAI released a browser. The race to mine browsing habbits for context is on.

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.

Nish Tahir @nishtahir.com · 11d

Don't interpret the study as Adopt AI now, there are drawbacks. It should be a guide to think about where it fits in your org

Nish Tahir @nishtahir.com · 11d

• Developer trust is the barrier to greater usage - Establish an AI usage policy, give people opportunity to experiment but don't force them to use it.
• There are mixed signals coming from other sources/survey - METR stands out and raises questions about the credibility of self reported data

Nish Tahir @nishtahir.com · 11d

TL;DR
• Mostly self reported data.
• 90% of devs use AI at work.
• Simply adopting tools fix your problems
• AI adoption is a force multiplier - if your delivery practice is good, you'll likely see some benefit. If it's bad, it'll likely make everything worse.

Nish Tahir @nishtahir.com · 11d

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 ...

Nish Tahir @nishtahir.com · 12d

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.

Nish Tahir @nishtahir.com · 14d

Alternatively get a Mac Mini or studio. Both expensive but fine options.

Nish Tahir @nishtahir.com · 14d

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...

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.

Nish Tahir @nishtahir.com · 15d

Overall it seems interesting and useful and something open source tools will likely replicate quickly. There's already an issue in opencode github.com/sst/opencode...

Support for "Skills" · Issue #3235 · sst/opencode

Yesterday Claude Code released a new feature called "Skills", that basically allows the user to define prompts, that are lazily loaded by the Agent when needed. https://www.anthropic.com/news/skill...

Nish Tahir @nishtahir.com · 15d

Skills are generally a markdown file called SKILL.md. There are premade examples in a repo here github.com/anthropics/s...

GitHub - anthropics/skills: Public repository for Skills

Public repository for Skills. Contribute to anthropics/skills development by creating an account on GitHub.

Nish Tahir @nishtahir.com · 15d

MCP focuses on trying to provide tools to the LLM that it can optionally use. These are more like smaller manuals the LM can reference to improve behavior on certain tasks.

You can have a skill on parsing PDFs as an example - Use pdf kit, ignore table of content, include page numbers - for example

Nish Tahir @nishtahir.com · 15d

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

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

Nish Tahir @nishtahir.com · 17d

Now that i'm looking at the numbers against the AMD AI MAX+ 395 platform, it feels like this should have been half the price.

Nish Tahir @nishtahir.com · 17d

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...

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

Nish Tahir @nishtahir.com · 17d

I haven't found a compiled apples to apples against the DGX spark yet but looks to get ~49.40 tps on GPT-OSS 120b mxfp4 www.reddit.com/r/LocalLLaMA.... It's presumably a more capable platform for dev work based on review comments.

Mushoz's comment on "MINISFORUM MS-S1 Max AI PC features AMD Strix Halo, 80 Gbps USB, 10 Gb LAN, and PCie x16 - Liliputing"

Explore this conversation and more from the LocalLLaMA community

Nish Tahir @nishtahir.com · 17d

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...

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