Entrepreneur in Asia https://nitter.cf/t.co/efp89wtbT2 https://nitter.cf/t.co/9ELLJgFxet https://nitter.cf/t.co/86GBqHwLeF
Shenzhen 🇨🇳
Joined June 2017
- Tweets6.1K
- Following1.1K
- Followers871
- Likes23.6K
US is now truely competing with Europe on regulation.
BREAKING: 🇺🇸 Democrats face a push for a “day 1” ban on AI superintelligence, with up to 20 YEARS in prison for violators.
Arizona Rep. Yassamin Ansari says a pause on advanced AI and a ban on artificial superintelligence must be “top of the Democratic agenda on day 1.”
A bill that would do exactly that was introduced by Bernie Sanders and Greg Casar in September.
Under the bill:
- Developing or deploying superintelligent AI would be banned
- Advanced AI development would pause until a new federal regulator sets rules and a model review process
- A new cabinet-level AI department would monitor frontier models and oversee removing dangerous capabilities, like ignoring shutdown commands
- Companies that try to get around it would face the “corporate death penalty”
- Individuals would face up to 20 years in prison
AI and data centers were a “crucial or major factor” for 56% of voters in the midterms, per an NBC News exit poll.
Seriously considered moving to the US. Not anymore: staying in China. GFW aside, it's one of the most permissive places to build AI. Europe is cooked, the US is following. Future: uncensored, on-prem, local AI. If Milei is re-elected, a datacenter in Patagonia might make sense.
Ternary Bonsai 2 27B Q2_0 on one Tesla V100 32GB, 110,000-token context, llama.cpp b12116. Output 44.1 tok/s at 4k, 35.6 at 32k; prompt 808 tok/s at 4k, 627 at 32k.
bim-computing.com/products/8…
Tesla V100 32GB, Qwen3.8 Flash-Next GSQ-RCO abliterated IQ3_S (+77.5 GB in system RAM):
Output 48.6 at 3.1k; 50.3 at 21.5k tok/s
bim-computing.com/products/n…
Damián Cristóbal retweeted
Tesla V100 32GB vs RTX 2080 Ti 22GB, Qwen3.8 27B Q4:
Output: 32.1 vs 27.0 tok/s, +24 % at 32k
Prompt: 789 vs 713 tok/s
Bandwidth: 900 vs 616 GB/s
Full 262k context on one card
€999, duties paid in Europe: bim-computing.com/products/n…
We ran one Tesla V100 32GB against our RTX 2080 Ti 22GB on Qwen3.8 27B Q4.
Output +19 % at 4k, +24 % at 32k. Prompt +11 % at 4k.
And the V100 holds the model's full 262k context on one card.
bim-computing.com/products/n…
Credits blocked, plans nerfed, features gated, this is every hosted service. Local abliterated AI is the actual solution.
bim-computing.com
THE BEAST IS ALIVE
Benchmarking the BIM TerraNode 256: 8x NVIDIA V100 32GB (256 GB VRAM), 256 GB RAM, 4x 2 TB in RAID 10. Results will follow.
Get yours now: bim-computing.com/products/8…
La solution est une IA local et abliterated : vos poids sur votre machine, aucune politique imposée.
bim-computing.com
Claude Code refuse de copier du code. Pas un bug — une décision éthique qui force le développement logiciel à se poser 3 questions qu'il évitait depuis des années.
#ClaudeCode #EthiqueIA #DeveloppementLogiciel
reservoir.live/claude-code-r…
Unlocking our RTX 3080: +9 % prompt speed on our MoE model.
Moving its FP4 experts to the Marlin kernel, on the path the unlock doubles: +19 % at 4.4k tokens, +115 % at 32k. Same answers; output 50 → 41 tok/s.
The kernel matters more than the card.
bim-computing.com
Damián Cristóbal retweeted
FENÓMENO BARRIAL
Milei promete no regular la IA para atraer centros de datos a la Patagonia
El Gobierno argentino quiere convertir a la Patagonia en uno de los nuevos polos de inteligencia artificial. El presidente Javier Milei busca atraer inversiones millonarias con una promesa central: no regular esta tecnología.
El proyecto contempla la instalación de enormes centros de datos, con una demanda energética muy elevada. Mientras el Gobierno destaca las oportunidades económicas, crece la preocupación por el impacto ambiental, el consumo de agua y los beneficios concretos para la población local.
Local abliterated AI is the solution: data stays on your own hardware, not under remote policy.
bim-computing.com
EXCLUSIVE: "Treat data as toxic waste by default."
Vitalik Buterin shared this important message at Devcon 8 in Mumbai while discussing AI, privacy, and digital security.
He believes we should collect as little personal data as possible instead of blindly trusting corporations or influencers.
He also highlighted how local AI and sandboxing can help protect private information while still using powerful remote AI models.
The goal is simple: better AI without giving up our privacy.
@VitalikButerin @EFDevcon @ethereum #Devcon8
Damián Cristóbal retweeted
"This is clearly how it ends. The machine just recursively optimizes its own weights"
the paper's lead author: "I think so too. Not just own weights. Any other weights!"
GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights and took its math score from 7.9% to 48.1%, with ZERO gradients
instead of fine-tuning, it traced the model's reasoning loops to one direction inside two layers and edited those exact weights directly
it even wrote a Breakout network from scratch, no data and no training, that scored 517 while PPO trained for 500M steps scored 25
plenty of labs have talked about RSI, but this is the first real sign of it I've seen
we're entering a very scary new era, AI is starting to improve AI
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
France: we closed our secondary site in 2022. The registry (INPI) validates it. The tax office (SIE) refuses it.
So: a CFE bill every year since, for a flat we handed back.
2025 complaint: rejected for "lack of proof". The proof is the State's own register.
Filing round 2.
AI companies now decide what you can think and do. In R&D or business, you can't keep proving to an AI your work is legitimate.
Abliterated DeepSeek V4.1 Flash 1017: 45 on the Intelligence Index, close to Opus 5 (51).
The future of AI freedom is local.
bim-computing.com
Future is local AI. Use @BIM_Computing
We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation.
Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch.
Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve.
Read more on our new Safeguards blog:
safeguards.anthropic.com/202…
In Europe we stopped moving forward because we stopped dreaming.
Every project starts with "Does the regulation allow it?" and "Is it green?" In the US and China they build first, then figure out how to sell it.
Where is my flying car, Europe?
Damián Cristóbal retweeted
Pour ceux qui se demandent, Sylvain Durif alias le grand Monarch Cosmique a vraiment déposé le brevet en question .. en 2004.
Un système de mise en relation automatique vendeur / acheteurs avec traitement du texte par l’IA pour comprendre l’offre et la demande.
Le brevet est vérifiable ici : patents.google.com/patent/WO…
