NVIDIA Shrinks DeepSeek V4.1 Flash by 75 GiB With Almost No Quality Loss
NVIDIA shipped an NVFP4 quantized DeepSeek V4.1 Flash that trims 75 GB off disk while matching the original on six benchmarks.
AI/ML news, top picks, and generated innovation digests.
90742 matching items
NVIDIA shipped an NVFP4 quantized DeepSeek V4.1 Flash that trims 75 GB off disk while matching the original on six benchmarks.
A refusal-stripped build of Qwen3.5-4B shrinks to 2.7GB, keeps benchmark scores intact, and runs offline on an 8GB laptop.
A study by Vals AI tested 25 AI models for alignment with Chinese government positions and found surprising results. Chinese models respond to political questions less uniformly than expected, and Western models occasionally adopt official Beijing positions too. The language of the prompt had an especially strong effect on the political slant of responses. The article Chinese AI models are politically complex, not just government mouthpieces, study finds appeared first on The Decoder .
Nigeria’s technology market is usually described by its problems, from unreliable infrastructure to a volatile economy. Dotun Daniel Adekunle, OPay’s chief operating officer and chief technology officer, argues that the country’s most durable companies are the ones built to survive its worst days. “Stubbornness is a business model in Africa,” Adekunle said in a keynote […]
From marble runs to advent calendars, these are the best Star Wars gifts for fans of all ages.
Moving back home from Los Angeles meant more time with family. When my dad suddenly passed, that meant even more.
Training against a monitor does not necessarily make a model more honest. Baker et al. (2025) showed that penalizing suspicious chain-of-thought can make models hide reward hacking rather than stop it. Cundy and Gleave (2025) found that the outcome depends on exploration, monitor accuracy, and KL regularization. More recently, The Obfuscation Atlas (2026) distinguished two ways models evade monitoring: changing their internal representations ( obfuscated activations ) and producing deceptive responses that the monitor fails to recognize ( obfuscated policies ). But when does RL learn to exploit a monitor's blind spots instead of stopping the unwanted behavior? Does the model need to develop a new evasive strategy during RL, or is it enough that this strategy already exists, even with a very low probability? I investigate this using small transformers and Qwen3-1.7B in controlled reward-hacking tasks. I find that making deceptive hacks sufficiently rare before RL can prevent monitor evasion, even without honesty training. A one-time edit to the model's output weights reproduces the protection of honesty training, here truthful supervised fine-tuning (Truthful SFT). Moreover, the activation probe does not appear to become worse at recognizing deceptive hacks during RL; it largely fails to recognize them even before RL begins. Experimental setup Toy setup I trained small transformers (4 layers, width 128) to add four numbers modulo 113. Each response is a narration token, three…
OhMyGame is an open-source desktop studio where an AI agent drafts the design doc, generates art, writes code, and playtests the game for you.
Pakistan takes charge of G24 developing nations bloc
This November on Netflix, watch the hit movie "Obsession," new release "The Mosquito Bowl," and classics "Cliffhanger," and "Weekend at Bernie's."
Iran ramps up its threats and does not exclude pre-emptive military action, but its officials have nevertheless noted positive steps in mediated negotiations with the US ahead of the decisive midterm elections. The Gulf countries remain "braced for the worst."
OpenAI is now capping the number of paid ChatGPT accounts per device and payment card. Users who hit the limit can't start new subscriptions, and affected accounts can't be used on that device. Logging out doesn't automatically free up a spot. The policy will likely hit users who've been juggling multiple cheap subscriptions to maximize usage, pushing them toward pricier options like buying extra credits. The article OpenAI is steering ChatGPT users toward more paid usage appeared first on The Decoder .
Africa’s creators have the talent and the audience. What holds them back is the infrastructure that turns attention into income.
Scenario Labs open-sourced 70+ Agent Skills that teach Claude, Cursor, and Codex to generate production-ready images, video, audio, and 3D through MCP.
Using artificial intelligence to build even smarter systems will make artificial general intelligence (AGI) more achievable, though it would need to be done safely, leading Chinese computer scientist Tang Jie said at an event in Shanghai on Saturday. The remarks underscore how AI companies in the US and China are both chasing the holy grail of recursive self-improvement (RSI), where increasingly powerful AI systems also become better at improving themselves in an ever-accelerating...
The Chromagraph Camera app costs $13 and processes iPhone photos with the same color math Rian Johnson’s cinematographer, Steve Yedlin, uses for film.
You may not have noticed everything Google’s free software suite offers—especially a web app called Google Vids.
"Appointing the new judges lies within the power of the Sejm," Commissioner Michael McGrath told POLITICO.
Anthropic's IPO filing laid out several ways in which it plans to give back. We asked three philanthropy gurus to assess its efforts.
The Villa Vie Odyssey kept circling one corner of it while finances worsened, services deteriorated, and cash calls mounted.
In this Sunday edition of Business Insider Today, we're talking about best practices for running (or watching) the NYC marathon.
As prices escalate on gaming laptops, there’s really only one affordable option still around: the Lenovo LOQ 15, which offers a better performance and display than its competitors.
These portable heaters will heat a room quickly, quietly, and safely.
A Google-led research team built StreamVibe, a video agent designed to watch people cook or assemble furniture and speak up only when help is needed. A benchmark shows that none of the tested models reliably pick the right moment in longer videos, as errors compound and response quality drops sharply. The article Google wants video AI to understand what people need as they work appeared first on The Decoder .
The vote puts together two candidates with sharply contrasting views on how Bulgaria should engage with Russia and the war in Ukraine moving forward.
딥시크가 여러 코딩 작업을 자동으로 관리하는 \'슈퍼 에이전트\'를 공개하며 앤트로픽의 \'클로드 코드\'와 오픈AI의 \'코덱스\' 추격에 나섰다. 지난 5월 전담팀 구성 이후 약 6개월 만에 자체 개발 도구와 팀 단위 업무 자동화 기술까지 선보였지만, 아직 연구용 프로토타입으로 실제 제품 경쟁력은 검증되지 않았다.딥시크 하네스 개발팀은 31일(현지시간) 공식 위챗 계정을 통해 \'슈퍼 에이전트 설계 탐구: 더 자율적인 에이전트 시스템을 향해\'라는 기술 문서를 발표했다. 이와 함께 새로운 구조를 구현한 \'DSH 슈퍼에이전트(dsh-supera
Catch up with the most important stories from around Europe and beyond this November 1st, 2026 - latest news, breaking news, World, Business, Entertainment, Politics, Culture, Travel.
A student-loan borrower was told he earned Public Service Loan Forgiveness, only to have his payments reinstated. Now he's a declarant in a lawsuit.
Founder Rachel Woods explains how her 8-person company uses AI agents to automate work and avoid 4 new hires. Here's her playbook.
With a number of anti-data-center candidates poised to win on Tuesday, New York provides a glimpse at the tensions around actually having to do something about the digital infrastructure.
Value-focused retailers are picking up market share and traditional chains are sliding as shoppers grapple with rising costs.
The private equity giant just hosted its biennial Apollo Opportunity Foundation summit for nonprofit leaders.
The intervention by the former British chancellor comes ahead of a UK-EU summit later this month.
These top picks from brands like Ember, Owala, and Hydro Flask are built to last, keeping your drinks hot and your mornings smooth.
I spent $333.11 to register for the 2026 New York City Marathon. I didn't bat an eye at checkout — but not because the purchase wasn't significant.
Outsourcing planning to AI can lead people to surrender their own critical thinking, which can have potentially deadly results in places like the backcountry.
The "stopover" hack, which about a dozen airlines offer, let me visit a bucket-list country without having to book another trip across the Atlantic.