"zero data retention" isn't a vendor-wide promise, it's an endpoint setting
openai needs separate approval for it, store:false alone isn't enough. anthropic keeps flagged content up to 2 years
worth checking which endpoints you're actually on
marktechpost.com/2026/11/17/…
FastFrames-L on 1x NVIDIA V100 32 GB: FastFrames-L produced 3/3 clean 5.06 s clips, but needed 31.3 GB VRAM and 277 s per clip on a 32 GB V100. Verdict: partly works. bim-computing.com
@GooglePlay@Google@sundarpichai@AnthropicAI
Absolutely DISGUSTED with the way Google Play treats paying customers.
I purchased a Claude subscription through Google Play and subsequently had my Claude account banned. I can no longer access the service I paid for.
I turned to Google Play for a refund, and guess what? REFUSED.
So Google is happy to take my money, but when something goes wrong and I can no longer use the service, you simply wash your hands of the problem?
This is beyond disappointing. It’s an absolute disgrace.
Your customer support is shockingly poor, your refund process is frustrating and unfair, and your complete lack of accountability is unacceptable.
Why should customers bear the financial loss for a service they can no longer access?
I am extremely frustrated and deeply disappointed with Google Play.
I DEMAND a proper review of my case and a full refund.
Is this really the standard of customer service Google stands for?
Co-operative AI Network (CO-AI):
A Federated, Community-Owned Artificial Intelligence Infrastructure for India. Collective ownership. Shared computing. Open technology.
India’s co-operative movement has a long history of people pooling resources to own and govern things that would be difficult for any one person to hold alone. Sugar factories in Maharashtra and dairy societies like @DairyAmul showed how farmers could collectively own processing and marketing infrastructure, share the benefits, and decide how it runs.
Artificial intelligence (AI) now faces a similar access problem. Open-weight models have made capable AI software available for download. The practical barrier for many individuals, professionals, and smaller organizations is the cost of the computing hardware, power, cooling, and day-to-day operations needed to run it.
The basic idea is straightforward. People who cannot individually own AI infrastructure should be able to collectively own it. Independent primary co-operative societies can own and operate local computing nodes. A federation can connect them for shared standards, resource discovery, and capacity exchange while each society keeps control of its own assets and members.
This is not proposed as just another AI service provider. It is a model in which communities participate as owners of the infrastructure that delivers the services. A concept paper on a Co-operative AI Network explores this approach, beginning with a pilot in Maharashtra.Discussion and feedback are welcome.
Disclaimer: This White Paper presents an original idea and conceptual framework by myself. AI tools were used strictly as supporting aids for language refinement and editorial polishing. Image generated using AI.
Cc: @grok for review...
UNCENSORED AI THAT SEES YOUR IMAGES AND RUNS ON ONE GPU
huihui abliterated qwen 3.5 35b a3b, a free multimodal moe that only fires 3b params at a time, so it runs fast on a single gpu
how to run it:
1. the official model
huggingface.co/huihui-ai/Hui…
2. ready-to-run ggufs
huggingface.co/mradermacher/…
3. 24gb gpu, download this one
Huihui-Qwen3.5-35B-A3B-abliterated.Q4_K_M.gguf
4. 16gb gpu, use this instead (12.9 gb)
Huihui-Qwen3.5-35B-A3B-abliterated.Q2_K.gguf
5. for images, also grab
Huihui-Qwen3.5-35B-A3B-abliterated.mmproj-Q8_0.gguf
6. load it in lm studio, ollama or llama.cpp
abliterated models can still refuse now and then, and they can lose a little smarts
Google now keeps some Gemini chats for around two months — even with history turned off. You can opt out of training, but not out of retention. Convenience or surveillance? Where's your line on AI privacy?
Haben Sie sich gefragt, warum eine KI, die auf Ihrem eigenen Rechner läuft, mehr wert ist als jede Cloud-Lösung? Mittwoch, 2. Dezember, 18:00 Uhr: Hermes – Lokale KI und eigene KI-Agenten auf deinen Geräten.
Eine KI, die auf Ihrem Gerät läuft, gehört Ihnen. Ihre Daten bleiben bei Ihnen. Kein Server in Virginia entscheidet, was mit Ihren Eingaben passiert.
Lokale KI-Modelle sind technisch anspruchsvoller. Aber sie lösen das Grundproblem jeder Cloud-KI: Sie wissen nicht, wer mitliest.
Genau das macht sie für Unternehmen relevant, die sensible Daten verarbeiten.
Im KI Learning Lab zeige ich, wie lokale Modelle funktionieren, wann sie Cloud-Lösungen überlegen sind und was Sie dafür brauchen. Anmeldung:
Mehr Infos:
+1 Grok refused me one something much less sinister than reverse engineering a game
Its a pussy now sadly
I wanna see grok win, but its losing on all fronts
ChatGPT @OpenAI can't code properly, missing details, asking again and again to fix it
Claude @AnthropicAI banned and paid no refund
No edge to use SI anymore, cancelled all SI subscriptions and no trading/ no vibecoding and taking a long break from here
Im tired of this shit
Gemma 4 on 1x NVIDIA V100 32 GB: Gemma 4 12B Q4_0 ran locally on 8.2 GB VRAM with all three quality prompts correct: arithmetic, JSON extraction and constrained writing. Verdict: works. bim-computing.com
A slow window by volume, an instructive one by composition: nine items, every single one a listing on Modelers, the model hub serving Huawei's Ascend ecosystem, and almost none of them new models. This is what infrastructure accumulation looks like — ports, distills, repackagings — and it raises the only question that matters here: which hardware stacks can actually run the open weights everyone keeps publishing.
The piece with real engineering weight is the FLUX.1 ONNX export. Black Forest Labs' 12-billion-parameter text-to-image model, packaged as FLUX.1-dev in ONNX under AMD's namespace, drops the native PyTorch dependency and runs today on ONNX Runtime and DirectML, with a documented conversion path toward Huawei's Ascend NPUs via CANN's ATC tool. Note the second-order effect: portability work published by one vendor hands a free win to rival silicon too. Format neutrality is the rare ecosystem play where everyone keeps the gains.
Three of the nine items are one recipe in different hats. The featured one is openthoughts-114k-no-special-template, a 7-billion-parameter model built on Alibaba's Qwen2.5-7B-Instruct that distills DeepSeek-R1's chain-of-thought style into a compact, permissively licensed package — specifically a chat-template ablation from the OpenThoughts research series. A second listing covers the same no-special-template experiment under the mlfoundations-dev banner; a third, from Monday evening, applies the identical OpenThoughts-114k trace set — roughly 114,000 reasoning traces distilled from DeepSeek-R1 — to a Qwen2.5-Math base. The framings differ (R1 capability arriving on Ascend, a lightweight reasoner drop, a distillation study), but the substance is one dataset and three fine-tunes. Publishing the stripped-template control is genuinely useful science. It is also a fair measure of how much 'new open reasoner' news is dataset recombination.
OpenGVLab doubles up on the hub. Husky-13B-delta is a unified agent built on Meta's Llama 2 13B, tuned for numerical, tabular and knowledge-based reasoning, with a pipeline that runs query understanding, task decomposition, action planning and solution synthesis, plus built-in tool use including web search and Python execution. The unified-agent thesis is the right one — a single model that decides when to reason and which tools to call beats a patchwork of specialists — though building on a Llama 2 base in late 2026 means the scaffolding is current and the foundation is not. Alongside it sits InternImage-B, a roughly 97-million-parameter vision backbone trained on ImageNet-1K at 224x224, permissively licensed, serving as both standalone classifier and feature extractor for detection and segmentation. Unglamorous, useful, and exactly what a hardware ecosystem needs: proven backbones that run where the customers' silicon actually is.
The bottom tier is where hub listing culture shows its seams. Two items cover one developer's lfm2-scenarios — a fine-tune of Liquid AI's LFM2 family, published with an almost empty model card, then repackaged as GGUF for consumer-grade hardware. Edge-friendly packaging is fine; a 'scenario-focused generator' whose documentation is nearly blank is a file upload dressed as a release. The same author's lfm2-physics repeats the pattern: Apache-2.0 license, a science label, and no stated objective, scope, or evaluation target. A permissive license makes a model easy to redistribute. It does not make it worth running.
The real story this window is infrastructural: between AMD's ONNX packaging and a hub steadily filling with Ascend-ready ports, the live contest in open AI is which runtimes can host the good models. The models are trending commodity. The runtimes are the moat.
Strata engine on 1x NVIDIA V100 32 GB: Strata 0.1.60 runs on V100 but reads prompts 11-39% slower than 0.1.59 and loads 58 s longer, so we keep 0.1.59. Verdict: partly works. bim-computing.com
32 GB of VRAM, two prices. 2x Tesla V100 16GB (ours): €998 delivered, €31.2/GB. RTX 5090 32GB: €9,300 today (LDLC), €290.6/GB. Decode holds up too, figures in the image. bim-computing.com
The smaller open-weight Qwen4 models will be out before the end of November, and the main one will be the 27B tier that can run on a single GPU.
It will be more usable than the last local 27B, especially for coding agents and long-context work. But general intelligence will only move forward in small steps: it won’t be a mini Max, and it won’t catch the frontier flagships. 4-bit will still be in the 16 to 19GB range, and how much usable context is left after quantization is what will actually decide the experience.
Tesla V100, liquid-cooled edition. 240 mm all-in-one cooler, quiet at the full 250 W. HBM2 900 GB/s.
32 GB: 1,099 EUR delivered. 16 GB: 599 EUR delivered.
Official NVIDIA datacenter driver, Ubuntu and Windows.
bim-computing.com/products/n…