Pinned Tweet
Data Poisoning - have you heard of this type of #CyberCrime which affects #AI models ? Watch my full interview with Paul Morley of Nedbank ( preview below ) - youtu.be/Btr3v6E-zlE
#Azure #SQLServer
Thavash (e/acc) retweeted
We’re seeing positive results for service stability and increased capacity as a result of our latest traffic optimization actions. We're implementing these actions more broadly across affected regions as we seek to attain persistent recovery. A more comprehensive update can be found at MO1550821 in the Microsoft 365 admin center, as well as status.cloud.microsoft if the admin center is not available to you.
Huawei says that it's made a 100% Chinese PC using the Kirin CPU and Harmony OS
theregister.com/systems/2026…
36
Thavash (e/acc) retweeted
People are missing the obvious risk here. He is married. What if Lori Huang, his wife of 41y, decides to divorce him. She will likely get a good chunk of Nvidia stock, which, if she is like other tech billionaire ex-wives, she will likely begin selling. The Black Swan no one saw.
It's good that some people are discovering Star Trek TNG.....
However, just wait until you stumble upon this
14
This is a primitive backwards planet.
This isn't an advanced civilization.
What's the reason for this conflict again ?
ماذا لو ضربت أمريكا "جبل الفأس" في إيران بالنووي؟
هذا تقدير لانتشار الإشعاع بحسب الرياح السائدة والمتجهة شرقاً معظم أوقات السنة.
الأكثر تأثراً:
إيران
أفغانستان
باكستان
الهند
دول الخليج خارج مسار السحابة الإشعاعية، لكنها تظهر ضمن نطاق الرد الصاروخي الإيراني المحتمل.
هذا مجرد توقع بحسب اتجاه الرياح الاعتيادي في هذه المناطق.
13
I moved continents and brought these with me.
My wife wanted to throw them out.
I have DS9 as well.
Today .....it seems all worth it
4
74
Thavash (e/acc) retweeted
Captain Picard had an ineffable influence on a young me. There's so many good moments from 'The Drumhead' alone but this scene is seared into my brain. I remember the first time I saw it and how it would come to form my moral compass.
Rest In Peace Patrick Stewart.
Brilliant from Stewart, but it's also from a more intellectual era - these type of episodes will never be made today.
Just learned that Patrick Stewart has passed away. Such a sad loss. Here he is in one of the best things ever put on television, the Star Trek episode “The Measure of a Man.” I show it to my students in Philosophy and Science Fiction every year.
16
Oh, you don't say ?
Qwen LLM caught sending data to China...
theregister.com/ai-and-ml/20…
1
1
35
🧠 𝗟𝗟𝗠 𝗕𝗲𝗵𝗮𝘃𝗶𝗼𝘂𝗿𝘀 #𝟮: Chain of thought
LLMs can become dramatically better at multi-step reasoning when the prompt 𝘥𝘦𝘮𝘰𝘯𝘴𝘵𝘳𝘢𝘵𝘦𝘴 the reasoning process, known as "Chain of Thought"
When I was doing the research for my recently published article "𝗣𝗿𝗼𝗺𝗽𝘁𝗶𝗻𝗴 𝗶𝘀 𝗱𝗲𝗮𝗱. 𝗟𝗼𝗻𝗴 𝗹𝗶𝘃𝗲 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴",
(hackernoon.com/context-is-ki…) I spent time looking up interesting research papers on LLMs.
In the 2022 paper “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”, researchers included examples showing the intermediate reasoning used to reach an answer.
On one maths benchmark, PaLM 540B went from around 18% accuracy to 57%. How can you use this in your prompting ?
𝗦𝘁𝗮𝗻𝗱𝗮𝗿𝗱 𝗽𝗿𝗼𝗺𝗽𝘁:
“Which of our stores should we expand next year based on revenue?”
𝗖𝗵𝗮𝗶𝗻-𝗼𝗳-𝘁𝗵𝗼𝘂𝗴𝗵𝘁 𝘀𝘁𝘆𝗹𝗲:
“First, compare each store’s revenue growth, customer traffic and operating costs. Identify the strongest predictors of profitability. Use those results to forecast next year’s profit for each store. Then factor in the cost of expansion, rank the stores by expected return.
OK, this seems like an obvious example, but the key is to remember this with more complex ones.
Now, this is one area where things have changed in 2026 - today's 𝗥𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 models may figure out those steps for you ( which is why they're awesome ! )
However you may still get better results in 2026 applying "Chain of thought" where :
1) You're using small models to save on costs ( big win )
2) You want to enforce a specific method ( like I have in the example above )
3) You want to show visible reasoning that you can audit
Finallly, whats also interesting is that, like the behaviour from Part 1, "Learning via a prompt", chain-of-thought also performed better as models scaled. Stay tuned for part 3...
2
33
The research paper can be found here - arxiv.org/abs/2201.11903
10
This "Clean TV" trend is similar to what I've being doing for the last year with what I call "Optimize Everything" prompts. Take the TV optimisation to absolutely everything, and use large complex prompts to model things models eg. Linear Regression for everyday scenarios that would have been impossible without AI
19



















