iiinova — where your digital self keeps evolving

lukerm @zl-labs.bsky.social · 24/02/2026

IBM suffers worst day (13% drop in stock price) in 25 years after Anthropic touts COBOL AI tool 📉 seekingalpha.com/article/4873... IBM Corporation suffered its largest one-day drop since the dot-com bubble, down over 20% YTD, prompting a downgrade from Buy to Hold.

seekingalpha.com

100

lukerm @zl-labs.bsky.social · 18/02/2026

From Simon Willison's blog: "These days I’m writing significantly more code via Claude Code than by typing text into a text editor myself. I’m actually getting more coding work done on my phone than on my laptop, thanks to the Claude Code interface in Anthropic’s Claude iPhone app.

100

lukerm @zl-labs.bsky.social · 13/02/2026

Claude Opus 4.6, and a range of other ultra-intelligent models, are now available for use in Parallellm! Try them here: www.parallellm.com Or try Opus 4.6 vs 4.5 to see the differences, and see whether the hype is really all it's cracked up to be! #ai #llm #parallellm

000

lukerm @zl-labs.bsky.social · 08/01/2026

Former GitHub CEO Nat Friedman's website - nat.org - is refreshingly simple: just plain text, bullet points, and hyperlinks. It's an interesting contrast to GitHub itself, which is one of the most complex popular platforms on the web! #github #homepage #simple

Image shows a very simple black and white homepage at nat.org. The title is "Nat Friedman", followed by simple concise bullet points containing tidbits about himself and what he believes. There are hyperlinks to several external websites, such as Wiki pages to companies he co-founded.

lukerm @zl-labs.bsky.social · 06/01/2026

"Anytime we see Claude do something incorrectly we add it to the CLAUDE.md, so Claude knows not to do it next time." venturebeat.com/technology/t... The inventor of Claude Code reveals his strategy - it's genius! #ai #claude #code

000

lukerm @zl-labs.bsky.social · 20/12/2025

I've been experimenting with Agentic AI for automated Quality Assurance (QA) for websites, in this case Parallellm. It works where traditional methods fail. Read more about it on the blog reel 👇 zl-labs.tech/post/2025-11... #agentic #ai #llm #parallellm

zl-labs.tech

AI Agents for Automated Quality Assurance

Utilizing LLMs for Automated Workflows

lukerm @zl-labs.bsky.social · 19/12/2025

You can now play Pick-A-Partridge on Reddit! 🎄 It's a festive game of memory that anyone can play. www.reddit.com/r/pick_a_par... This little side project turned into a bit of a "thing". Have fun!🎄 #reddit #devvit #christmas

lukerm @zl-labs.bsky.social · 07/11/2025

Moonshot AI just released Kimi K2 ... with Thinking. Now, it's available for use in Parallellm! parallellm.com This huge model has made a splash, achieving SOTA on several agentic benchmarks, such as Humanity's Last Exam and 𝜏²-Bench Telecom. Another great one added 👍 #ai #llm #parallellm

lukerm @zl-labs.bsky.social · 05/11/2025

A few weeks ago, AWS went down. Hard. Here's Fireship's funny post-mortem video of how it happened: www.youtube.com/watch?v=HqL0... A misconfigured DNS brings down half the internet 🤦‍♂️ #aws #fireship #fail

US-EAST-1 is humanity’s weakest link…

YouTube video by Fireship

020

lukerm @zl-labs.bsky.social · 28/10/2025

DeltaDB. "A collaborative workspace where humans and AI agents work together across a range of time scales, with every insight preserved and linked to the code forever." zed.dev/blog/sequoia... This could be interesting... #ai #git #zed

000

lukerm @zl-labs.bsky.social · 17/10/2025

I made demo video to walk you through how to use Parallellm www.youtube.com/watch?v=tUmH... #ai #parallellm

YouTube video by Parallellm

Parallellm: Product Demo

lukerm @zl-labs.bsky.social · 17/10/2025

How to use n8n, with a Fireship twist. I've fast-forwarded to the demo. www.youtube.com/watch?v=bS9R... It's even open source! #workflow #oss

n8n will change your life as a developer...

YouTube video by Fireship

lukerm @zl-labs.bsky.social · 10/10/2025

Tuesday brought lots of excitement as pi-Python (3.14) - or "Pithon", if you will - was released. It's all about the colour: syntax highlighting straight in the REPL! Check out the difference between versions 3.13 and 3.14! #python #pithon

lukerm @zl-labs.bsky.social · 02/10/2025

Beta launch of Parallellm 🚀 parallellm.com I'm absolutely thrilled to announce the launch of Parallellm in Beta - supercharge your AI conversations in parallel! 🫵 Beta testers needed! Try it out the platform for FREE, and share your feedback! 🙏

parallellm.com

Parallellm

Large Language Models, Supercharged in Parallel

100

lukerm @zl-labs.bsky.social · 01/10/2025

Sora 2 is out, and here's the masterplan: "The majority of users ... should feel that their life is better for using Sora that it would have been if they hadn’t. If that’s not the case, we will make significant changes..." blog.samaltman.com/sora-2#:~:te... Let's see how that plays out ... #ai

blog.samaltman.com

Sora 2

We are launching a new app called Sora. This is a combination of a new model called Sora 2, and a new product that makes it easy to create, share, and view videos. This feels to many of us like...

000

lukerm @zl-labs.bsky.social · 19/09/2025

A great post-mortem analysis of Windsurf's demise: ethanding.substack.com/p/windsurf-g... " Claude Code charges $5/day ($150/month) as a "good proxy for how much those 'should' cost," ... Cursor charges $20/month. Windsurf charges $15/month. They're selling dollar bills for quarters. " #ai

ethanding.substack.com

windsurf gets margin called

what the windsurf sale means for the ai coding ecosystem

100

lukerm @zl-labs.bsky.social · 05/09/2025

Then and now ... the state of the code in 2025. Times are changing! (Comic from Andrew Ng's newsletter) #ai #agentic

Left (2022): manager asks employee to code the "Fizzbuzz" program.
Right (2025): manager asks employee to build an entire e-commerce platform by Friday.

Machine Learning @machinelearning.bsky.social · 05/09/2025

Reposted by lukerm

Build a fully-local voice assistant via LangGraph + MCP servers with no subscriptions, no cloud, just fast, on-device smarts that actually work. The future of personal AI is private and modular. #AI #VoiceAssistant #LangGraph #MCP #Privacy #ModularAI #LocalAI

towardsdatascience.com

Using LangGraph and MCP Servers to Create My Own Voice Assistant | Towards Data Science

Built over 14 days, all locally run, no API keys, cloud services, or subscription fees.

033

lukerm @zl-labs.bsky.social · 25/08/2025

A great talk by Andrej Karpathy on the "software update" in light of LLMs and friends. We're currently in the 1960s era of computing when the computers were massive and centralized. Well worth a watch! www.youtube.com/watch?v=LCEm... #ai #llm

Andrej Karpathy: Software Is Changing (Again)

YouTube video by Y Combinator

000

lukerm @zl-labs.bsky.social · 18/07/2025

AI-Powered XKCD Comic Finder! Read more about it on my blog 👇 bit.ly/zl-labs-xkcd... Demo available here: xkcd-finder.com #AI #MachineLearning #SemanticSearch #RAG #xkcd

Retrieving XKCD Comics with Semantic Search and RAG

XKCD Finder

021

lukerm @zl-labs.bsky.social · 20/06/2025

This map of GitHub is awesome! I guess we're somewhere around the Learning Cortex 🧠 Link: anvaka.github.io/map-of-github #ai #machinelearning #github

A virtual map showing interconnected projects on GitHub, grouped by region or domain. It has a cool blue tone with different regions having various shades of blue. The map specifically highlights the "Learning Cortex" which is where the author's work resides.

lukerm @zl-labs.bsky.social · 11/06/2025

An amazing, thoughtful and philosophical question about AI, asked by a young person most likely to be affected by the outcome 👏 "If you start talking to AI, if AI was in a human body, and talks and acts like a human, and you become friends with it, would it count as social interacting?"

100

lukerm @zl-labs.bsky.social · 06/05/2025

OpenAI is changing its company structure to a Public Benefit Corporation. Jumped or pushed? Here's an interesting discussion on the history of this development. www.youtube.com/watch?v=GwQV... #generative #ai #openai

BREAKING: OpenAI ABANDONS Plan To Go For Profit

YouTube video by Matthew Berman

000

lukerm @zl-labs.bsky.social · 07/04/2025

The Parallellm Pump is a new developer power tool for comparing responses from LLMs in parallel. Read more about it on my blog here 👇 bit.ly/zl-labs-parallellm-pump #genai #llm #ai #claude #chatgpt #deepseek #ml

A power tool for comparing LLM responses

Parallellm Pump

000

lukerm @zl-labs.bsky.social · 03/04/2025

[Sound on] I'm speaking with the V2V MoshiVis model in the video (NOT the purple stick figure 🙂) MoshiVis holds much promise, but there are evidently a few kinks to be ironed out. Still, a great technology in the making! #ai #ml #genai #transformers #moshi

lukerm @zl-labs.bsky.social · 25/03/2025

Anthropic's Claude is great at coding, but does make mistakes. Part 2 (see image) suggests that you can use a lambda function to combine columns during a pandas aggregation. This is confabulation! As demo'd in the image. The rest is correct. Go carefully! #GenAI #llm #ai #claude #pandas

Multi-panel image. Left panel shows Claude giving detailed answer to the question: "In pandas, when you group by, can you apply a function involving two of the columns in the group?" Claude then suggests three different methods. The remaining two panels look at suggestion 2 of its response. The central panel demonstrates that you can use the .agg function to apply column-wise functions, but the third panels shows that an error is thrown on attempting to aggregate across two columns with a lambda function, as suggested by part 2 in Claude's response.

lukerm @zl-labs.bsky.social · 06/03/2025

Does Claude 3.7 really "enable" bioweapons acquisition? An interesting RCT by Anthropic studies whether access to their LLM makes this possible - thankfully it doesn't! 😅 Participants made too many critical errors.

$The image shows data visualizations from a "Bioweapons Acquisition Uplift Trial" with two side-by-side box plots. Left chart - "Full Score": Shows fraction scores (0-1.0) across different groups divided into "Sepal" and "Internal" categories. The groups include various combinations like "Sepal + Google," "Sepal + Claude 3.7 Sonnet," and "Internal + Claude 3.7 Sonnet" variations. A horizontal dashed line at approximately 0.8 and a dotted line around 0.5 mark reference thresholds. Right chart - "Number of Critical Failures": Shows the count of failures (0-8) for the same groups. The "Internal + Google" group shows the highest median number of failures (around 5-6), while several Claude-based systems show lower failure counts (mostly 2-3).$

100

lukerm @zl-labs.bsky.social · 19/02/2025

Sharing my experiences of running a 32B DeepSeek-R1 LLM directly on AWS in a blog post. Lots of interesting tips and tricks in there. Read it here 👇 bit.ly/zl-labs-deep... Let me know your thoughts! #ml #genai #LLM #ai #machinelearning #pytorch

The screen recording shows an example of the DeepSeek model generating its thinking process in real-time (sped up 8x) whilst responding to the following humdinger of a question:

“Do any two people in London have the same number of hairs on their heads?”

The panels on the right show the resource utilization within the cloud instance during generation. The top-right pane monitors the usage of the four GPUs by constantly polling gpustat. You can see that, after loading the model into GPU memory, the cores heat up and their usage fluctuates wildly between 15-40%. The bottom-right shows the CPU usage, mostly engaged during loading the model into GPU memory, and thereafter idle apart from one core.

001

lukerm @zl-labs.bsky.social · 14/02/2025

Wait! This token will improve reasoning performance. A recent paper "s1: Simple Test-Time Scaling" found a fascinating insight: literally using the extra token "Wait" forced Qwen2.5-32B-Instruct) to continue "thinking", allowing it to double-check it's previous answer ...

Scatter plot titled 'Competition Math (AIME24)' with accuracy (%) on the y-axis and thinking time (tokens) on the x-axis. It shows thinking time (measured in output tokens) is positively correlated with average correctness accuracy. It highlights that extending reasoning by appending 'Wait' improves accuracy by encouraging longer thinking times (6000–8192 tokens).

100

lukerm @zl-labs.bsky.social · 07/02/2025

DeepSeek API is back online! 🥳 But you can't top up your balance yet 🤦 Until that changes, we're stuck with running it locally! #DeepSeek #AI #genAI #LLM #whatthe

Image showing that the DeepSeek API service is up and running, but with a large orange banner across the top stating that it's not possible to top up your funds. This makes the service effectively useless to new users.

000

Reposted by lukerm

Gabriel Archanjo @garchanjo.com · 05/02/2025

"Xi" + " Jin" + "ping" = "Let`s talk about something else". #DeepSeek

109812

lukerm @zl-labs.bsky.social · 28/01/2025

It turns out you can just initiate your .gitignore files by using boilerplates from Github's github repo! 🤯 With commits from over 13 yrs ago, I'm wondering how I didn't know this sooner. Schooled. Enjoy this tidbit if it's new. #git #gitignore #CodeGold #schoolday

The main page of the gitignore project by github, showing .gitignore templates for various programming languages.

000

lukerm @zl-labs.bsky.social · 23/01/2025

ChatGPT is down. Worldwide productivity grinds to a halt. People stare blankly into their screens not knowing what to do. How do I write this? Poorly written social-media posts ensue. Sorry, had to write this one with my own, uncalibrated cortex. Darn, run out of #AGI #LLM #ChatGPT #ChatGPTdown

ChatGPT showing significant outages on Downdetector.

030

lukerm @zl-labs.bsky.social · 19/01/2025

DeepSeek-V3 is an open-weights LLM with comparable or enhanced performance against GPT-4o across a range of technical benchmarks (MMLU, Math 500, ...). Built with just $5.6m training compute 🤯 Read more about it: www.deeplearning.ai/the-batch/de... #AI #genAI #LLM

deeplearning.ai

DeepSeek-V3 Redefines LLM Performance and Cost Efficiency

A new model from Hangzhou upstart DeepSeek delivers outstanding performance and may change the equation for training costs.

021

lukerm @zl-labs.bsky.social · 15/01/2025

Here is a worrying trend of banks withdrawing from the NZBA (green action group) in response to upcoming transfer of power: www.theguardian.com/business/202... On a lighter note, you can keep up-to-date on the USA's position on climate here: does-usa-support-paris-agreement.com #netzero #climate

does-usa-support-paris-agreement.com

🇺🇸 Does The USA Support The Paris Agreement?

Up-to-date status on the USA's wavering support on protecting the climate

000

lukerm @zl-labs.bsky.social · 09/01/2025

Pop-ups. Pop-ups and cookie banners stand between us and Artificial General Intelligence (AGI). In 'Benchmarking LLM Agents on Consequential Real World Tasks', the authors study the (in)ability of a cohort of agents when completing sequential route and mundane tasks. 1/4

lukerm @zl-labs.bsky.social · 24/12/2024

1/11 The latest blog post from ZL Labs was entitled: "Let it Segment: A Gift From SAM": bit.ly/zl-labs-sam2 Coming up with this title was a collaboration between me and a popular LLM. I thought it would be instructive to publish the transcript documenting process. So here it is ☃️

Let it Segment: A Gift from SAM

Automatically segmenting festive images using SAM models

100

lukerm @zl-labs.bsky.social · 20/12/2024

One final blog post from me for the year - this time with a festive feel 🎄⛄ bit.ly/zl-labs-sam2 What else should be segmented? Where could this go next? Let me know your thoughts. Happy Holidays! #ml #ai #machinelearning #pytorch #python #holidays #snowman #feelingfestive #ZLlabs

Let it Segment: A Gift from SAM

Automatically segmenting festive images using SAM models

000

lukerm @zl-labs.bsky.social · 19/12/2024

My favourite AI failchievement of 2024 has got to be the epic video of ChatGPT trying to make ASCII art of the word 'HONDA': www.reddit.com/r/ChatGPT/co... Closely followed by SORA failing to catch the fish: www.youtube.com/watch?v=2jKV... Let's see what progress 2025 brings! #ML #AI #genAI