Stuart Gray
@sgray.bsky.social
He/Him. AI Wrangler. Web Geek. F1 Fan. All views my own.
🤖 AI, LLMs, GenAI, NLP
🐍 Python Dev
🚀 Indie Hacker
🎮 Game Dev, ProcGen, Unity, C#
🏎️ F1 Fan
🇬🇧 UK Based
🦣 mastodonapp.uk/@StuartGray
✖️ x.com/StuartGray (inactive)
🤖 AI, LLMs, GenAI, NLP
🐍 Python Dev
🚀 Indie Hacker
🎮 Game Dev, ProcGen, Unity, C#
🏎️ F1 Fan
🇬🇧 UK Based
🦣 mastodonapp.uk/@StuartGray
✖️ x.com/StuartGray (inactive)
Pinned
Stuart Gray
@sgray.bsky.social
· Mar 27
I welcome any genuine civil discussion, challenge, or critique.
However, if you strongly disagree with a post to the point you're unable to refrain from insults, rude or unthinking replies then please, save us both a lot of time and block me now - because I will block you.
However, if you strongly disagree with a post to the point you're unable to refrain from insults, rude or unthinking replies then please, save us both a lot of time and block me now - because I will block you.
Reposted by Stuart Gray
Yet another reminder that when it comes to data privacy, Microsoft cannot be trusted -- and, worse, its management is an eager ally of the fascists who run our government.
https://techcrunch.com/2026/01/23/microsoft-gave-fbi-a-set-of-bitlocker-encryption-keys-to-unlock-suspects-laptops-reports/
https://techcrunch.com/2026/01/23/microsoft-gave-fbi-a-set-of-bitlocker-encryption-keys-to-unlock-suspects-laptops-reports/
Microsoft gave FBI a set of BitLocker encryption keys to unlock suspects’ laptops: reports
The FBI served Microsoft a warrant requesting encryption recovery keys to decrypt the hard drives of people involved in an alleged fraud case in Guam.
techcrunch.com
January 24, 2026 at 7:13 AM
Yet another reminder that when it comes to data privacy, Microsoft cannot be trusted -- and, worse, its management is an eager ally of the fascists who run our government.
https://techcrunch.com/2026/01/23/microsoft-gave-fbi-a-set-of-bitlocker-encryption-keys-to-unlock-suspects-laptops-reports/
https://techcrunch.com/2026/01/23/microsoft-gave-fbi-a-set-of-bitlocker-encryption-keys-to-unlock-suspects-laptops-reports/
Thinking about this some more, I really need to see the specific wording in the legislation because on the face of it this would ban 16 & 17 year olds from working remotely for organisations.
A nice way to inadvertently increase youth unemployment 🤦♂️
A nice way to inadvertently increase youth unemployment 🤦♂️
Not sure how I missed this, but it seems the UK’s House of Lords has decided to ignore King Canute tidal advice and voted to ban VPNs for under 18s 🤦♂️
Good luck with that!
www.techradar.com/vpn/vpn-priv...
Good luck with that!
www.techradar.com/vpn/vpn-priv...
UK government targets VPNs in online safety consultation as Lords vote for ban
Technology Secretary Liz Kendall promises "evidence-led" approach to keeping children safe online
www.techradar.com
January 23, 2026 at 11:59 PM
Thinking about this some more, I really need to see the specific wording in the legislation because on the face of it this would ban 16 & 17 year olds from working remotely for organisations.
A nice way to inadvertently increase youth unemployment 🤦♂️
A nice way to inadvertently increase youth unemployment 🤦♂️
Not sure how I missed this, but it seems the UK’s House of Lords has decided to ignore King Canute tidal advice and voted to ban VPNs for under 18s 🤦♂️
Good luck with that!
www.techradar.com/vpn/vpn-priv...
Good luck with that!
www.techradar.com/vpn/vpn-priv...
UK government targets VPNs in online safety consultation as Lords vote for ban
Technology Secretary Liz Kendall promises "evidence-led" approach to keeping children safe online
www.techradar.com
January 23, 2026 at 10:42 PM
Not sure how I missed this, but it seems the UK’s House of Lords has decided to ignore King Canute tidal advice and voted to ban VPNs for under 18s 🤦♂️
Good luck with that!
www.techradar.com/vpn/vpn-priv...
Good luck with that!
www.techradar.com/vpn/vpn-priv...
Reposted by Stuart Gray
From a thread about LLMs in a classroom. Awesome thread in general, but also this cool summary of an interesting paper: it's old, but relevant, as it's more about philosophy and general trends than about who in particlar is getting replaced in year of the lord 20XX
bhaven.org/uploads/3/4/...
bhaven.org/uploads/3/4/...
January 23, 2026 at 8:16 AM
From a thread about LLMs in a classroom. Awesome thread in general, but also this cool summary of an interesting paper: it's old, but relevant, as it's more about philosophy and general trends than about who in particlar is getting replaced in year of the lord 20XX
bhaven.org/uploads/3/4/...
bhaven.org/uploads/3/4/...
Reposted by Stuart Gray
Fascinating pic from @cip.org. People show surprisingly high levels of trust in AI systems. Low trust in AI companies. Full report here static1.squarespace.com/static/631d0...
January 22, 2026 at 8:25 AM
Fascinating pic from @cip.org. People show surprisingly high levels of trust in AI systems. Low trust in AI companies. Full report here static1.squarespace.com/static/631d0...
Reposted by Stuart Gray
Open source, locally runnable, realtime on consumer hardware (define consumer) world model?! Crazy!
foo
bar
over.world
January 20, 2026 at 9:19 PM
Open source, locally runnable, realtime on consumer hardware (define consumer) world model?! Crazy!
Reposted by Stuart Gray
Reposted by Stuart Gray
Reposted by Stuart Gray
Media may not have noticed that Trump’s Board of Peace Charter was drafted so that the US need NOT ratify and become a party.Trump’s current status as US President is a necesary stepping stone but his unitary rule in the new international organisation would be personal, permanent and even hereditary
January 21, 2026 at 6:31 AM
Media may not have noticed that Trump’s Board of Peace Charter was drafted so that the US need NOT ratify and become a party.Trump’s current status as US President is a necesary stepping stone but his unitary rule in the new international organisation would be personal, permanent and even hereditary
Reposted by Stuart Gray
Davos, Switzerland, the World Economic Forum.
January 20, 2026 at 10:16 PM
Davos, Switzerland, the World Economic Forum.
Reposted by Stuart Gray
I've often joked that as faculty I program in a high-level language called "graduate student". Having tried out Claude Code this morning, I (i) feel extremely at home, (ii) am realizing that research-by-graduate-student is perhaps the original vibe-coding. 1/2
January 8, 2026 at 12:24 PM
I've often joked that as faculty I program in a high-level language called "graduate student". Having tried out Claude Code this morning, I (i) feel extremely at home, (ii) am realizing that research-by-graduate-student is perhaps the original vibe-coding. 1/2
A bit late to the party, but I just learned about Qwen 3 Coder REAP 25B
A pruned version of the 30B model with near zero quality loss, so you can run a higher quant or more context (GGUFs are available)
huggingface.co/cerebras/Qwe...
A pruned version of the 30B model with near zero quality loss, so you can run a higher quant or more context (GGUFs are available)
huggingface.co/cerebras/Qwe...
cerebras/Qwen3-Coder-REAP-25B-A3B · Hugging Face
This model was obtained by uniformly pruning 20% of experts in Qwen3-Coder-30B-A3B-Instruct using the REAP method.
huggingface.co
January 20, 2026 at 12:47 PM
A bit late to the party, but I just learned about Qwen 3 Coder REAP 25B
A pruned version of the 30B model with near zero quality loss, so you can run a higher quant or more context (GGUFs are available)
huggingface.co/cerebras/Qwe...
A pruned version of the 30B model with near zero quality loss, so you can run a higher quant or more context (GGUFs are available)
huggingface.co/cerebras/Qwe...
An interesting read & finding, with an equally interesting dilemma for model providers;
They can reduce sycophancy & AI psychosis relatively easily & cheaply.
BUT
It comes at the cost of also eliminating role play, a significant use of LLMs and source of revenue.
They can reduce sycophancy & AI psychosis relatively easily & cheaply.
BUT
It comes at the cost of also eliminating role play, a significant use of LLMs and source of revenue.
This is interesting for a lot of reasons, including explaining how model personality drift happens (& a way to mitigate that) as well as more exploration into the “Assistant” the key personality of basically every AI you work with, but which is not well understood. www.anthropic.com/research/ass...
The assistant axis: situating and stabilizing the character of large language models
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
www.anthropic.com
January 20, 2026 at 12:28 PM
An interesting read & finding, with an equally interesting dilemma for model providers;
They can reduce sycophancy & AI psychosis relatively easily & cheaply.
BUT
It comes at the cost of also eliminating role play, a significant use of LLMs and source of revenue.
They can reduce sycophancy & AI psychosis relatively easily & cheaply.
BUT
It comes at the cost of also eliminating role play, a significant use of LLMs and source of revenue.
Reposted by Stuart Gray
Who is the Big Tobacco of today?
In new work, we find 50% of high profile social media papers are connected to big tech through funding, collaboration and employment. Most connections aren't disclosed. @jbakcoleman.bsky.social @jevinwest.bsky.social @carlbergstrom.com 1
arxiv.org/abs/2601.11507
In new work, we find 50% of high profile social media papers are connected to big tech through funding, collaboration and employment. Most connections aren't disclosed. @jbakcoleman.bsky.social @jevinwest.bsky.social @carlbergstrom.com 1
arxiv.org/abs/2601.11507
Industry Influence in High-Profile Social Media Research
To what extent is social media research independent from industry influence? Leveraging openly available data, we show that half of the research published in top journals has disclosable ties to indus...
arxiv.org
January 19, 2026 at 6:51 PM
Who is the Big Tobacco of today?
In new work, we find 50% of high profile social media papers are connected to big tech through funding, collaboration and employment. Most connections aren't disclosed. @jbakcoleman.bsky.social @jevinwest.bsky.social @carlbergstrom.com 1
arxiv.org/abs/2601.11507
In new work, we find 50% of high profile social media papers are connected to big tech through funding, collaboration and employment. Most connections aren't disclosed. @jbakcoleman.bsky.social @jevinwest.bsky.social @carlbergstrom.com 1
arxiv.org/abs/2601.11507
Reposted by Stuart Gray
Great post, and an excellent example of the Bitter Lesson of AI: generalized models are always over time going to beat specialized models. substack.com/home/post/p-...
Gemini 3 Solves Handwriting Recognition and it’s a Bitter Lesson
Testing shows that Gemini 3 has effectively solved handwriting on English texts, one of the oldest problems in AI, achieving expert human levels of performance.
substack.com
January 19, 2026 at 10:29 PM
Great post, and an excellent example of the Bitter Lesson of AI: generalized models are always over time going to beat specialized models. substack.com/home/post/p-...
Ooh, llama.cpp server now supports the Anthropic Messages API, allowing you to use Claude-compatible clients with locally-running models, including Claude Code.
huggingface.co/blog/ggml-or...
huggingface.co/blog/ggml-or...
New in llama.cpp: Anthropic Messages API
A Blog post by ggml.ai on Hugging Face
huggingface.co
January 19, 2026 at 10:18 PM
Ooh, llama.cpp server now supports the Anthropic Messages API, allowing you to use Claude-compatible clients with locally-running models, including Claude Code.
huggingface.co/blog/ggml-or...
huggingface.co/blog/ggml-or...
Reposted by Stuart Gray
Insights from Ben Affleck on AI:
• AI can help write scenes but can't create full movies.
• Job loss fears are overblown because adoption of new tech is slow; it's hype for startup valuations.
• ChatGPT v5 is ~25% better but costs 4x.
• Users actually preferred v4's sycophancy for companionship.
• AI can help write scenes but can't create full movies.
• Job loss fears are overblown because adoption of new tech is slow; it's hype for startup valuations.
• ChatGPT v5 is ~25% better but costs 4x.
• Users actually preferred v4's sycophancy for companionship.
January 18, 2026 at 5:32 AM
Insights from Ben Affleck on AI:
• AI can help write scenes but can't create full movies.
• Job loss fears are overblown because adoption of new tech is slow; it's hype for startup valuations.
• ChatGPT v5 is ~25% better but costs 4x.
• Users actually preferred v4's sycophancy for companionship.
• AI can help write scenes but can't create full movies.
• Job loss fears are overblown because adoption of new tech is slow; it's hype for startup valuations.
• ChatGPT v5 is ~25% better but costs 4x.
• Users actually preferred v4's sycophancy for companionship.
Reposted by Stuart Gray
my father teaches computer science and he's describing this variance too: the kids who were flunking before still are now, but the kids who love this stuff are making the best projects that college has ever seen by leaps and bounds
January 19, 2026 at 8:57 PM
my father teaches computer science and he's describing this variance too: the kids who were flunking before still are now, but the kids who love this stuff are making the best projects that college has ever seen by leaps and bounds
Reposted by Stuart Gray
Everyone has heard that AI complicates traditional forms of student assessment.
What you haven't heard is that AI also makes it possible for motivated students to design & execute quite interesting projects, so some profs (cough) are now regularly running 4-6 independent studies. Sustainable? 🤷♂️
What you haven't heard is that AI also makes it possible for motivated students to design & execute quite interesting projects, so some profs (cough) are now regularly running 4-6 independent studies. Sustainable? 🤷♂️
January 19, 2026 at 8:35 PM
Everyone has heard that AI complicates traditional forms of student assessment.
What you haven't heard is that AI also makes it possible for motivated students to design & execute quite interesting projects, so some profs (cough) are now regularly running 4-6 independent studies. Sustainable? 🤷♂️
What you haven't heard is that AI also makes it possible for motivated students to design & execute quite interesting projects, so some profs (cough) are now regularly running 4-6 independent studies. Sustainable? 🤷♂️
Reposted by Stuart Gray
After making dozens of prototypes with AI coding agents, here's what I think: human developers are still essential
In fact, instead of taking jobs, this new class of software tools may make workers busier than ever. I wrote about my experience for Ars Technica:
arstechnica.com/information-...
In fact, instead of taking jobs, this new class of software tools may make workers busier than ever. I wrote about my experience for Ars Technica:
arstechnica.com/information-...
10 things I learned from burning myself out with AI coding agents
Opinion: As software power tools, AI agents may make people busier than ever before.
arstechnica.com
January 19, 2026 at 2:08 PM
After making dozens of prototypes with AI coding agents, here's what I think: human developers are still essential
In fact, instead of taking jobs, this new class of software tools may make workers busier than ever. I wrote about my experience for Ars Technica:
arstechnica.com/information-...
In fact, instead of taking jobs, this new class of software tools may make workers busier than ever. I wrote about my experience for Ars Technica:
arstechnica.com/information-...
Reposted by Stuart Gray
An interesting bit of history:
In 1916, Denmark sold the Danish West Indies to the US. You may know of them as the US Virgin Islands. As part of this sale, the US recognized Danish authority over Greenland.
The text of the complete, ratified agreement is here: www.govinfo.gov/content/pkg/...
In 1916, Denmark sold the Danish West Indies to the US. You may know of them as the US Virgin Islands. As part of this sale, the US recognized Danish authority over Greenland.
The text of the complete, ratified agreement is here: www.govinfo.gov/content/pkg/...
January 19, 2026 at 7:06 AM
An interesting bit of history:
In 1916, Denmark sold the Danish West Indies to the US. You may know of them as the US Virgin Islands. As part of this sale, the US recognized Danish authority over Greenland.
The text of the complete, ratified agreement is here: www.govinfo.gov/content/pkg/...
In 1916, Denmark sold the Danish West Indies to the US. You may know of them as the US Virgin Islands. As part of this sale, the US recognized Danish authority over Greenland.
The text of the complete, ratified agreement is here: www.govinfo.gov/content/pkg/...
Why is Trump kicking off about Greenland *now* specifically?
I can’t help wondering if it’s all related to the Nobel prize deadline?
There are two key blockers to peace in Ukraine, Putin & Europe.
Now Trump has leverage over Europe;
“I’ll settle with Putin or give me Greenland”, which is it?
I can’t help wondering if it’s all related to the Nobel prize deadline?
There are two key blockers to peace in Ukraine, Putin & Europe.
Now Trump has leverage over Europe;
“I’ll settle with Putin or give me Greenland”, which is it?
January 19, 2026 at 8:18 AM
Is there a term for “gambling” on how long you keep a vibe coding session going before actually trying to run the code?
I’m thinking “Vibe Roulette”?
Bonus points for how large/complex the app is, and how capable (or not) the model is at the task.
I’m thinking “Vibe Roulette”?
Bonus points for how large/complex the app is, and how capable (or not) the model is at the task.
January 18, 2026 at 8:52 PM
Is there a term for “gambling” on how long you keep a vibe coding session going before actually trying to run the code?
I’m thinking “Vibe Roulette”?
Bonus points for how large/complex the app is, and how capable (or not) the model is at the task.
I’m thinking “Vibe Roulette”?
Bonus points for how large/complex the app is, and how capable (or not) the model is at the task.
Reposted by Stuart Gray
As someone said the other day, one of the biggest problems with AI is not that you're expected to be more productive; it's that it triples the output of the biggest idiot you work with
January 17, 2026 at 11:21 AM
As someone said the other day, one of the biggest problems with AI is not that you're expected to be more productive; it's that it triples the output of the biggest idiot you work with
Reposted by Stuart Gray
