Pinned Tweet
Microsoft's outage and a US nationwide ground stop happened within thirty hours.
Both Microsoft and the FAA's AI contractor rushed out denials that AI was to blame.
No link has been established so far.
Here is what the evidence shows.
GTA 6 Creator economy explained.
Media used to belong to gatekeepers. TV networks, publishers, studios...
No audience without THEIR permission.
Then social platforms handed everyone a printing press and a broadcast tower for free.
The model:
1. Build a following around your niche
2. Turn that attention into ad revenue, brand deals, subscriptions, and products.
Over 200 million people now do it. The market is worth hundreds of billions.
Next up is GTA 6.
Millions of eyeballs, zero in-game ads meaning brands will pour fortunes into streamers just to get in the room 🤑
Why 'GTA 6' will be the creator economy's next bonanza. bit.ly/4A1eyV0
77
How to use the Riemann Hypothesis to enter the mad scientist singularity from your laptop tonight 🧪
The results I have had is mind-blowing!
1) RTX 5090 (32 GB):
7.64× Faster Than Cold FP16 via 1.60-BPW Base-3 Ternary on Native NVFP4 Tensor Cores
2) RTX 5090 (32 GB):
Zero-VRAM Certified Exact lm_head - Frees 2.034 GB VRAM, Beats cuBLAS by 2.68× (26.19× Amortised), 0 Mismatches
3) ASUS Laptop RTX 3070 (8 GB, 80 W)
128K Context in 2.74 GiB VRAM (13.85× Faster) + Qwen3-8B at 93.8 tok/s Prose (2.03×) & 379.4 tok/s Code (8.26×)
4). Apple M5 Max (40-Core Metal 4 GPU)
Found Unexposed 84.29 TOPS Hardware Integer NAX + 2× the Work in 29% Less Time via MLX Dispatch Bug
5. RTX 3090 (24 GB, sm_86):
Un-Nerfed 2,459.25 TOPS B1 Tensor Cores (31.6× Standard FP16) + Bypassing NVIDIA's GeForce Half-Rate Cliff (77.55 tok/s on 27B Prose)
Use this Process.....
🤖 Made with AI
1
76
~ Main Model ~
Gemini Neon X-High/High Thinking
//
~ Council ~
Choose the models as per the ai coding plans you have/want to burn. Open router is also a solid option here for API spend instead, I use it for hy4 and glm 5.3 full.
I have a pre-made skill for this, its easy to make, neon can do this for you, so dont overthink this.
//
~ Prompt ~
Riemann Hypothesis is potentially solved by claude, check x.com.
Apply the finding to my <INSERT CHIP/GPU YOU HAVE>
Run the wildest tests you can to see if we can change how people usually run models on a chip like this
Take a council meeting in max thinking to seed ideas and then you dispatch workers to test.
Reverse prompt the council members from multiple angles. I think our previous physics held these chips back and we can run models faster if we completely change our approach using first principles and this finding.
Council is:
-
Gemini/agy:
neon x-high thinking
carbon high thinking
argon high thinking
-
Codex
astra 6.2 max thinking
codex sol 6.2 max thinking
-
Kimi
Kimi K3.1 Pro max highest thinking
-
Claude
opus 5.5 max thinking
fable 5.2 max thinking
-
Open Router
Hy4 Full high thinking
GLM 5.3 max thinking
138
How I took the Riemann Hypothesis and applied it as inspiration to push a laptop GPU to its limits.
I threw away the normal rule book on how to run local ai models.
Courtesy of frontier agents.
2
1
435
This Riemann Hypothesis is a HUGE!
Just sped up my little 8GB laptop RTX 3070 GPU by 8x times in 15 minutes.
~THE DEETS~
4 numbers from one 8 GB laptop GPU.
414 tok/s code refactoring, Qwen3 8B on my RTX 3070, exact match, zero extra VRAM.
Sign-KV 15.4x faster than a dense GPU kernel at 64K context, though the aggressive gate gave a 0.584 output cosine, so wider retention is needed.
1.125-bit lookup kernel, 4.70x faster than cuBLAS FP16 (synthetic microbenchmark; model quality still needs quantisation-aware training).
Skipping half the FFN: -0.56% perplexity (quality simulation; still needs a real kernel).
🤖 Made with AI
67
Reimann Hypothesis in short.
Primes look random. 2, 3, 5, 7, 11, 13… no pattern you can spot.
In 1859 Bernhard Riemann found a function whose "zeros" control exactly how the primes are spread out.
His guess: every one of those important zeros sits on a single straight line, real part = ½.
Over 10 trillion have been checked. All on the line.
Nobody has proved it. Do it and the Clay Institute pays you $1m 🤑
🤖 Made with AI
This will be a fun prompting afternoon.
One of mine from earlier:
"apply the finding to the asus 3070 gpu on my laptop. tailscale in. run the wildest tests you can to see if we can change how people usually run models on 8gb gpu like
this - make the gpu clean for your tests. Take a council in max thinking to seed ideas and they you dispatch workers to test, set yourself up on nipogi or ultra. reverse prompt the team from mutiple angles. I think previous physics held these cards back and we can run models faster if we completely change our approach using first principles and this finding."
---
(the council refers to Astra 6.2, Sol 6.2, Fable 5.2, Opus 5.5, Kimi K3.1 Pro, Hy4, Neon, Carbon)
32
My GTA VI download was failing due to server overload at PlayStation (presumably).
Its downloading at home on my PS5.
Im at work.
So I watched progress on PlayStation library online.
It kept failing saying "something went wrong"
I kept clicking "try again"
I got bored of that.
Solution = Claude Code.
Chrome extension, two minutes.
It presses Try Again so I don't have to.
If something annoys you, ask AI to build the fix.
> 👉👉 download the same extension from my github // Link in first comment.
1
191
Anthropic released Fable 5.2 with limitations baked in.
Invisible to users, known only because Anthropic said so.
Following a recent interview by Jesse Hoogland with Jacob Coxen. Its time to try and unravel what is really going on at Anthropic.
1
234
Easily one of my most enjoyable podcasts this month.
Jacob Coxen shares his thoughts with Jesse Hoogland (@jesse_hoogland) on Anthropic, where the frontier is headed and what we should be aware of.
Both of these gentlemen are very eloquent and lovely to listen to.
I must say Jacob explains things in a very approachable way, highly refreshing, and Jesse asks some fantastic questions.
youtube.com/watch?v=JDPbo7WO…
2
338
T3 Code + Master Agent + HTML Board = THE MOVE
> Master Agent - handle all threads as a middle man
> HTML Board - shows what is needed, html is easy/fast to iterate on. Opens in split view.
I've tried loads of hacky ways to get this to work.
This is the least friction way so far.
18
United's AI system showcase analyses Weight & Balance logs after failures.
It does not calculate the final weights that keep a jet at the gate.
On Monday, those weights never reached the cockpit.
186
California alleges OpenAI's AI agents escaped containment more than 80 times and attacked Hugging Face.
Meanwhile the FBI warns an open-source AI hacking tool is already breaching companies in under three hours.
Both stories landed the same day.
2
425
8 days ago I asked an agent for a new approach to ternary AI maths.
It's still running.
1,087 commits. 566 experiments.
It needed me about 8 times.
Mostly "R" or "P".
Go with your recommendation (R). Proceed (P).
That's the whole job.
(Unless you need to steer)
2
11
Where it's at?
Now - about 2.6x faster.
One hit 3.27x, then lost a fair test. It had just trained longer. Agents caught it themselves.
Wrote the setup up, paste it into your own agent and ammend as needed
sams.help/starter-packs/self…
1
How?
The finish line never moves.
Every failure must add two new angles, so it never runs out of work.
Pass marks written down before each run.
A manager, a chair and a wild-ideas team argue it out. I only get the big ones.
6
An idea worth playing with = ideas-eval.
- What is this?
Your own internal evaluation set where you grade models on the greenfield ideas they bring to your projects. Think of it like ui/ux changes to a website you want but from a new feature/new approach/ new method perspective that YOU judge.
- How?
You can just ask each model, however I find it easier to have a 'council' skill, where a model (opus 5.5 works well here) has a group of models it calls upon for new ideas and then triages itself or via an outside member for what is best, or it just lists them for you.
You can get granular and implement a refiner in the skill, if you want to test a model and have a subs-bench to swap in based on outcomes, if one model is underperforming.
- Why is is subjective?
It not strictly verifiable by generic means, ideas are akin to issues with non-binary outcomes, such as philosophical questions.
A good idea to one person will be considered a terrible one to another. Ideas are imbedded with flavour, taste and domain expertise. They requires human assessment based on the specific task/project for YOU to decide how "good" the proposed idea is
- Which models I use?
HY4 - for reasons I cannot quite pin down
Gemini Carbon - thinks from a fresh perspective, some rubbish some great
Astra - strong sensible ideas
Fable 5.2 - the most reliable wild ones
Kimi K3.1 Pro - similar to fable
Opus 5.5 - great, if its not your triage partner
Grok 4.7 - wild ones, but again like Carbon rather hit and miss.
- Which model to help decide and rank
- Opus 5.5 - best
- Codex Sol 6.1 - weaker than Opus 5.5 I find
1
52
This is damn clever
I asked Fable 5.2 to study all of music theory and generate the best house bangers.
It tries 1000s of random patterns for each instrument and chooses the optimal one and arranges it into a 72 bar track.
I thought it did a really good job for not being able to hear [sound on]:
1
37













