LightNews — News, ranked by scholars

@wittywebhandle.bsky.social

Oh we're doing this one again?

This is IDENTICAL to what OpenAI and Anthropic were saying about DeepSeek V3 and R1

This is an admission from Western AI that their 'Moat' is SO WEAK that you can simply 'steal' a large portion of that value by training on its exhaust.

THAT'S NOT GOOD FOR YOU DARIO!

Anthropic accuses Chinese rival Alibaba of illicitly extracting AI capabilities

June 25, 2026 at 2:55 PM

The Register

@theregister.com

DeepSeek's not the only Chinese LLM maker OpenAI and pals have to worry about. Right, Alibaba?

Qwen 2.5 Max tops both DS V3 and GPT-4o, cloud giant claims Analysis The speed and efficiency at which DeepSeek claims to be training large language models (LLMs) competitive with America's best has been a reality check for Silicon Valley. However,…

dlvr.it

January 30, 2025 at 3:10 AM

Mario Zechner

@mariozechner.at

Siehe die Model-Beschreibung hier:
huggingface.co/openai/whisp...

Sowie das Paper zum Modell:
cdn.openai.com/papers/whisp...

openai/whisper-large-v3-turbo · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

October 1, 2025 at 9:53 AM

Nicole Hennig

@nic221.bsky.social

DeepSeek-V3 now runs at 20 tokens per second on Mac Studio, and that’s a nightmare for OpenAI https://venturebeat.com/ai/deepseek-v3-now-runs-at-20-tokens-per-second-on-mac-studio-and-thats-a-nightmare-for-openai/ #AI #DeepSeek

Text Shot: DeepSeek-V3-0324 employs a mixture-of-experts (MoE) architecture that fundamentally reimagines how large language models operate. Traditional models activate their entire parameter count for every task, but DeepSeek’s approach activates only about 37 billion of its 685 billion parameters during specific tasks.

This selective activation represents a paradigm shift in model efficiency. By activating only the most relevant “expert” parameters for each specific task, DeepSeek achieves performance comparable to much larger fully-activated models while drastically reducing computational demands.

March 27, 2025 at 2:18 AM

Ed Zitron

@edzitron.com

DeepSeek has upended the paradigm that building generative AI requires an endless supply of money and brand new NVIDIA GPUs, as well as the myth that OpenAI had some sort of magical unique selling point in reasoning. OpenAI's biggest product is now a commodity.

www.wheresyoured.at/deep-impact/

It’s a cynical, vulgar version of David and Goliath, where a tech startup backed by a shadowy Chinese hedge fund with $5.5 billion dollars under management is somehow the plucky upstart against the lumbering, lossy, oafish $150 billion startup backed by a public tech company with a market capitalization of $3.2 trillion.

DeepSeek's V3 model — which is comparable (and competitive with) both OpenAI's GPT 4o and Anthropic's Claude Sonnet 3.5 models (which has some reasoning features) — is 53 times cheaper to run when using the company’s own cloud services. And, as noted above, said model is effectively free for anyone to use — locally or in their own cloud instances — and can be taken by any commercial enterprise and turned into a product of their own, should they so desire.

In essence, DeepSeek — and I'll get into its background and the concerns people might have about its Chinese origins — released two models that perform competitively (and even beat) models from both OpenAI and Anthropic, undercut them in price, and made them open, undermining not just the economics of the biggest generative AI companies, but laying bare exactly how they work. That last point is particularly important when it comes to OpenAI's reasoning model, which specifically hid its chain of thought for fear of "unsafe thoughts" that might "manipulate the customer," then muttered under their breath that the actual reason was that it was a "competitive advantage."

And let's be completely clear: OpenAI's literal only competitive advantage against Meta and Anthropic was its "reasoning" models (o1 and o3, which is currently in research preview). Although I mentioned that Anthropic’s Claude Sonnet 3.5 model has some reasoning features, they’re comparatively more rudimentary than those in o1 and o3.

In an AI context, reasoning works by breaking down a prompt into a series of different steps with "considerations" of different approaches — effectively a Large Language Model checking its work as it goes, with no thinking involved, because these models do not "think" or "know" stuff. OpenAI rushed to launch its o1 reasoning model last year because, and I quote Fortune, Sam Altman was "eager to prove to potential investors in the company's latest funding round that OpenAI remains at the forefront of AI development." And, as I noted at the time, it was not particularly reliable, failing to accurately count the number of times the letter ‘r’ appeared in the word “strawberry.”

At this point, it's fairly obvious that OpenAI wasn’t anywhere near the “forefront of AI development,” and now that its competitive advantage is effectively gone, there are genuine doubts about what comes next for the company.

DeepSeek’s models actually exist, they work (at least, by the standards of hallucination-prone LLMs that don’t, at the risk of repeating myself, know anything in the true meaning of the word), they've been independently verified to be competitive in performance, and they are magnitudes cheaper in price than those from both hyperscalers (EG: Google's Gemini, Meta's Llama, Amazon Q, etc.) and from OpenAI and Anthropic.

DeepSeek's models don't require massive new data centers (they run on the GPUs currently used to run services like ChatGPT, and can even work on more austere hardware), nor do they require an endless supply of bigger, faster NVIDIA GPUs every year to progress. The entire AI bubble was inflated based on the premise that these models were simply impossible to build without burning massive amounts of cash, straining the power grid, and blowing past emissions goals, and that these were necessary costs to create "powerful AI."

Obviously, that wasn’t true. Now the markets are asking a very reasonable question: “did we just waste $200 billion?”

January 29, 2025 at 4:42 PM

Unofficial Hackaday Feed

@hackaday-feed.bsky.social

New Open Source DeepSeek V3 Language Model Making Waves

In the world of large language models (LLMs) there tend to be relatively few upsets ever since OpenAI barged onto the scene with its transformer-based GPT models a few years …read more

hackaday.com

January 27, 2025 at 9:30 PM

Simon Willison

@simon.fedi.simonwillison.net.ap.brid.gy

Blogged some notes on the new (still MIT licensed) Whisper Turbo model, quietly released by OpenAI yesterday

It’s both smaller and 8x faster than their previous Whisper Large https://simonwillison.net/2024/Oct/1/whisper-large-v3-turbo-model/

And you can run it on a Mac with “pip install […]

Original post on fedi.simonwillison.net

fedi.simonwillison.net

October 1, 2024 at 3:44 PM

shannon//cannon

@majorarlene.net

specifically, it uses Whisper 3:
huggingface.co/openai/whisp...
I don't care if it's "useful" or a "less bad" use of AI, it's still AI and I'm not gonna support it. couldn't have used any of that 500k pounds raised on Kickstarter for localisation?

openai/whisper-large-v3 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

October 14, 2025 at 6:34 AM

elvis

@eos.bsky.social

voyage-3 is claimed to outperform OpenAI's text-embedding-3-large by 7.55% and 3.82% respectively while offering smaller embedding dimensions (1024), longer context windows (32K tokens), and significantly lower costs (2.2x less than OpenAI v3 large, at $0.06 per 1M tokens).

November 24, 2024 at 6:05 PM

Papabear26

@pbr26.bsky.social

Sub vs. monster, what’s in your glass? v3. Tag from @cindyfcreations.bsky.social
#AiArtCommunity #digitalart #AiArt #aidigitalart #glass #remixratpack #surrealart #3D #design #INSTAGRAM #classic #cool #portrait #openai #AiCommunity #AiArtChallenge #SynthArt #CreativeAi #Ai #AiArtwork prompt in ALT

$A surreal, ultra-detailed close-up of a crystal-clear drinking glass filled with water, sitting on a reflective surface under dramatic lighting. Inside the glass, a miniature submarine—intricate, brass-riveted, with glowing portholes—navigates the confined space. From the depths of the water, a monstrous sea creature emerges impossibly large for the tiny environment: writhing tentacles pressing against the glass walls, sharp claws scraping the surface, bioluminescent eyes peering through the rippling water. The submarine’s searchlight beams refract and distort through the glass, scattering across the table in golden streaks, while bubbles and water currents swirl violently inside. The surreal juxtaposition of an everyday drinking glass containing a deep-sea battle creates a fantastical, dreamlike, and claustrophobic scene. Rendered in ultra-realism with cinematic lighting, hyper-detailed glass reflections, rippling water distortion, and monstrous textures filling the confined vessel.$

September 13, 2025 at 9:32 AM

Papabear26

@pbr26.bsky.social

Faithful Esmeralda v3.
#AiArtCommunity #digitalart #AiArt #aidigitalart #faithful #Esmeralda #wings #remixratpack #surrealart #3D #design #INSTAGRAM #classic #cool #portrait #openai #AiCommunity #AiArtChallenge #SynthArt #CreativeAi #Ai #AiArtwork prompt in ALT

Surreal haute couture fashion photography inspired by Esmeralda, a faithful, ethereal woman standing inside a vast Gothic cathedral that feels otherworldly rather than religious, wearing an emerald green couture gown with flowing silk layers, sculptural tailoring, and delicate gold embroidery, modern high-fashion silhouette, elegant and modest yet powerful pose, hands gently clasped at chest level, calm resolute expression. Large luminous angelic wings made of soft white feathers extend gracefully from her back, glowing subtly with pearlescent light. The environment is dreamlike and fantastical: floating mist at floor level, glowing firefly-like particles drifting through the air, faint celestial light rays, and a distant surreal castle silhouette framed within the cathedral arches. Cinematic editorial lighting with a balance of cool teal shadows and warm candle-like highlights, dramatic chiaroscuro, shallow depth of field, luxury fashion magazine aesthetic, ultra-detailed fabric textures, soft motion in the gown, magical realism tone, reverent yet mythic atmosphere, photorealistic fantasy, portrait orientation 9:16.

January 3, 2026 at 8:50 AM

Hacker News Top Stories

@hackernewsbot.bsky.social

32k context length text embedding models Discussion

voyage-3 & voyage-3-lite: A new generation of small yet mighty general-purpose embedding models

TL;DR – We are excited to announce voyage-3 and voyage-3-lite embedding models, advancing the frontier of retrieval quality, latency, and cost. voyage-3 outperforms OpenAI v3 large by 7.55% on aver…

blog.voyageai.com

November 24, 2024 at 4:00 AM

Ed Zitron

@edzitron.com

The AI bubble was inflated because we live in a high-information, low-processing society, where we trust investors, the media and analysts to be far more critical than they are, leading to AI getting credit for what it "could do" rather than what it does today.
www.wheresyoured.at/premium-the-...

How Did We Get Here?
So, as a little catchup, all of this bullshit started in November 2022 when ChatGPT launched, at a time when everybody was sort of realizing that there were no more hypergrowth ideas (what I call the Rot-Com Bubble). Large Language Models require thousands of GPUs (Graphics Processing Units) to both train models and run “inference” — the process through which a model creates an output like text or an image — and these GPUs are both power-intensive and create a lot of heat, requiring expensive cooling systems that are both extremely power- and water-hungry.

Back in 2022 (and 2023), people were convinced of the “Scaling Laws” of AI (based on a 2020 paper out of OpenAI, here’s a great piece by Charlie Meyer on the subject), that suggested that (to quote Meyer) there was “compelling evidence that increasing the size of large language models would increase their performance,” and models became larger by feeding them more training data.

The jump from GPT-3.5 to GPT-4 was huge, with dramatically more-nuanced outputs, and people believed that every subsequent jump would be as impressive.

It wasn’t. By November 2024, it became apparent that model developers had hit diminishing returns on model training, and were struggling to find new ways to improve them. Their only hope was “test-time compute” — giving models more compute power to calculate an output using reasoning models that take an input and make a plan of action that the model then executes, getting better outputs while spending a bunch more tokens, increasing the cost of inference dramatically.

But that didn’t really fix anything other than “how to make AI startups spend more money on compute,” and categorically failing to fix the only problem that mattered — that every single AI company is unprofitable because of the cost of running AI model-based services.

Since then, we’ve kind of been in a weird, eerie holding pattern. In January 2025, a Chinese hedge fund would release DeepSeek R1, a r…

DeepSeek was immediately memory-holed by every major AI company, none of whom decided to read the papers published by DeepSeek about how the model was trained, other than OpenAI vaguely suggesting it might sue the company because ChatGPT outputs were used to train DeepSeek R1 and v3.

Nevertheless, nobody in America seemed to bother to learn from DeepSeek, and instead set on a path to build even bigger, more huger models, such as GPT-4.5 ($400 million to train) and Claude Opus 4.

Otherwise, “progress” has been thin, as have new or meaningful products since GPT-4’s release in 2023, other than OpenAI’s awkward and expensive video generation app Sora, which I’ll get to later. Models have got “better” in some indeterminate way based on benchmarks that are either rigged in their favor or that they cheat on.

2025 was not the year of agents, as they get things wrong 70% of the time according to Carnegie Mellon University, and have a 40% fail rate on single step queries (literally asking a question and getting an answer) and a 65% fail-rate on multi-step queries (as in “go and look this up and then do something else”) according to a study by Salesforce.

In fact, Large Language Models are basically in the same place they were a year ago, except now everybody is anxious about the AI bubble, because nothing has really happened other than companies raising then burning hundreds of millions or billions of dollars.

November 14, 2025 at 7:00 PM

Gradient Flow

@gradientflow.com

Large Language Models and Foundation Model Futures: Key Trends for 2025 and Beyond
#AI #LLM #LLMs
gradientflow.com/foundation-m...

Foundation Models: What's Next for 2025 and Beyond - Gradient Flow

Recent releases from Google (Gemini 2.0 Flash), OpenAI (o1 & o3), and particularly DeepSeek (V3 & R1) have underscored the rapid pace of innovation in the foundation model space. This prompted me to c...

gradientflow.com

February 5, 2025 at 3:38 PM

matsuu

@matsuu.bsky.social

「Whisper Large V3 Turboは、Whisper Large V3と比較して約3.16倍高速に動作」つよい。ご視聴ありがとうございました問題対策いいね。わいわい

OpenAI Whisper Large V3 Turboモデルでリアルタイム文字起こしを試してみた | DevelopersIO

dev.classmethod.jp

November 9, 2024 at 9:30 AM

Rodolfo De Nadai

@rdenadai.dev

Chupa sociedade...

huggingface.co/openai/whisp...

openai/whisper-large-v3 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

June 18, 2025 at 4:52 PM

lunamoth

@lunamoth.bsky.social

FFmpeg 8.0, Whisper 지원 추가 | GeekNews

news.hada.io/topic?id=22508

이거 괜찮군요. 저는 Buzz 라는 앱 쓰다가 팟플레이어에서 Faster-Whisper-XXL, large-v3, 음성 강조 필터, VAD 필터 사용해서 음성 텍스트 변환 자막 만들어서 보고 있습니다

메타에서 Whisper 업데이트 버전 공개 해줬으면 좋겠군요

FFmpeg 8.0, Whisper 지원 추가 | GeekNews

FFmpeg 8.0 버전은 Whisper 음성 인식 모델 지원 기능을 공식적으로 추가함Whisper는 OpenAI가 개발한 오픈소스 음성 인식 모델로, 다양한 언어 자동 음성 변환에 사용됨이 기능 도입으로 비디오 및 오디오 처리 작업에서 직접 음성-텍스트 변환 워크플로우 자동화 가능성 증대개발자들과 미디어 자동화 분야에서 FFmpeg 활용성 및 효율성이 크

news.hada.io

August 14, 2025 at 10:09 AM

Doll

@dollspace.gay

No idea, but heres the model card huggingface.co/openai/whisp...

openai/whisper-large-v3 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

May 24, 2026 at 11:03 PM

kotobato

@kotobato.jp

書き起こしのモデルを OpenAI Large V2とGroq Large V3で比較したい気もするけど TeslaやStarbucks同様Groq自体を使いたくないわけで…

January 27, 2025 at 8:23 AM

LesPhaith

@lesphaith.bsky.social

Chinese AI startup DeepSeek has quietly released a new large language model that’s already sending ripples through the artificial intelligence industry
venturebeat.com/ai/deepseek-...

DeepSeek-V3 now runs at 20 tokens per second on Mac Studio, and that’s a nightmare for OpenAI

DeepSeek's free 685B-parameter AI model runs at 20 tokens/second on Apple's Mac Studio, outperforming Claude Sonnet while using just 200 watts, challenging OpenAI's cloud-dependent business model.

venturebeat.com

March 26, 2025 at 8:51 AM

Bob E Hayes

@bobehayes.bsky.social

China's DeepSeek releases #AI model upgrade, intensifies rivalry with OpenAI

"The new model, DeepSeek-V3-0324, was made available through AI development platform Hugging Face (@hf.co), marking the company's latest push to establish itself in the rapidly evolving AI market."

www.reuters.com/tech...

China's DeepSeek releases AI model upgrade, intensifies rivalry with OpenAI

Chinese artificial intelligence startup DeepSeek released a major upgrade to its V3 large language model, intensifying competition with U.S. tech leaders like OpenAI and Anthropic.

www.reuters.com

March 26, 2025 at 10:27 AM

Micha the DevOp

@michabbb.bsky.social

#VoyageAI introduces voyage-context-3, a contextualized chunk #embedding #llm that captures both chunk details and full document context 🔍 #ai

🔄 Outperforms #OpenAI-v3-large by 14.24% on chunk-level and 12.56% on document-level retrieval tasks

🧵👇