deepseek’s holding 幻方量化 is a quant company, many years already,super smart guys with top math background; happened to own a lot GPU for trading/mining purpose, and deepseek is their side project for squeezing those gpus

OpenAI's Deep Research is just a search+read+reasoning in a while-loop, right? unless i'm missing miss something, here is my replicate of it in nodejs, using gemini-flash and jina reader github.com/jina-ai/node-Deep…

Letter-dropping physics comparison: o3-mini vs. deepseek-r1 vs. claude-3.5 in one-shot - which is the best? Prompt: Create a JavaScript animation of falling letters with realistic physics. The letters should: * Appear randomly at the top of the screen with varying sizes * Fall under Earth's gravity (9.8 m/s²) * Have collision detection based on their actual letter shapes * Interact with other letters, ground, and screen boundaries * Have density properties similar to water * Dynamically adapt to screen size changes * Display on a dark background

not really. their holding quant company is highly profitable afaik and the founder does not need to market that - it’s already quite famous in Chinese quant/HFT community. i believe the founder just considers his quant business quite stable and has time and money to try something new and hence a new company deepseek since late 2023

When u really want to wrap up the week with deepseek-r1, and then @Alibaba_Qwen released Qwen-2.5-1M long-context model on Sunday evening.

perfect "open model" doesn’t exist—and it never will. every so-called "open" project has its boundaries, hidden corners, or legal booby-traps. it's often not bc the model devs are evil overlords hoarding secrets; but bc we’re all stuck in a world where ethics, lawsuits, corporate interests, personal convictions, and about a million other messy factors get tangled up in the conversation. so here's the thing: I’m just here for the ride. If you release a model, slap on a license, and let the rest of us tinker, break it, and have a blast—cool, I’m in.

so my reply went viral...let me add sth here. i know 幻方量化 high-flyer long time ago and even back in the late 2023 i heard people say they were running deepseek llm as side project bc of the leftover gpu. but nobody even in china takes them seriously. So it’s not that chinese ai teams r lean & great and can do such such great things; but it’s only deepseek lean & mean - chinese ai companies are just as fat and heavy on marketing just like their american counterpart. 2 things that make deepseek great: - the ceo is such a low-key guy, smart no ego & keep learning, never waste time on public exposure. - they spent years in quant - where the community values leverage & efficiency much more than headcount. And one person can and should manage 7 digits dollar portfolio without panic. so the lean & mean is deeply rooted in their culture.

Rationale是我司推出的一款专为管理者和决策者打造的分析成效工具，它集成了最新的GPT3.x和上下文学习(in-context learning)技术，能够快速生成Pros & Cons和SWOT分析报告，帮助管理者和个人做出明智的决策。rationale.jina.ai @mranti

colab.research.google.com/dr… Here is an updated notebook for training Phi-3-tiny (66M parameters) from scratch on @MSFTResearch TinyStories dataset using @huggingface SFTTrainer. This is inspired by @astar_research STLM proposal. And it just works! With 24GB VRAM you can do batch size of 64 with gradient accumulation at 2; probably can even do better. One can clearly see an improvement shortly after 1000 steps (which is ~6% of the full training data, so far from one epoch). Some remarks tho:

As everyone hails ChatGPT API, we had to speak up: our migration from davinci003 to gpt35-turbo actually made the generated content quality worse in many cases. While saving costs may be tempting, it's not worth sacrificing quality. Are we alone on this? #ChatGPT

Watch me talk about the neural search ecosystem from @JinaAI_. Since I founded the company in 2020, we have developed an #opensource product landscape for helping developers build deep learning-powered search applications. Follow me to know more! piped.video/watch?v=dvaj3u4E…

Looking at their src code, no 3p deps used. Claude went w/ React, others plain HTML+JS. Only Claude's ver has rotating free-fall letters vs others' upright-only fall.

Just dropped my DSPy intro slides from yesterday's presentation! blog-files.jina.ai/2024/04/D… They're designed with beginners in mind, featuring more visuals and less code to help you grasp the why and how of DSPy for prompt engineering. Also, I've cleared up some misunderstanding from my last blog post after watching an awesome presentation from @lateinteraction. Should've watched that sooner, but hey, better late than never!

also a quite kind low-key guy, he donated good amount to red cross or sth to help child disease study

deepseek ceo is not from chinese stanford (i.e tsinghua) no US education, not a phd student of some big prof., no google/msra work exp, too low-key; there r so many chinese guys meet those criterions, so why would china take this ceo srsly?

Excited to announce 🪩 𝗗𝗶𝘀𝗰𝗼𝗔𝗿𝘁: create compelling Disco Diffusion artworks in just one line! Radically easy, fully optimized for Google Colab free tier. github.com/jina-ai/discoart @multimodalart #creativeai #generativeart #opensource

🤯 Build a "Zoom" in 20 lines of code with Jina and #Python . Check how: github.com/jina-ai/jina-vide… A great showcase of how powerful, easy, and efficient Jina is even for real-time streaming services. We will call it "Joom"! #opensource #neuralsearch #multimodal @tiangolo

🇨🇳🇺🇸 are /acc. Steel sharpens steel; embracing competition, not dodging it, will make both stronger.

DALL·E Flow now has an official Docker image and requires only one GPU! Easy deployment! Thanks to the optimized diffusion step and @borisdayma new mega checkpoint, these are the fine artworks we get from DALL·E Flow! github.com/jina-ai/dalle-flo… @multimodalart

🎉 Unveiling 𝗝𝗶𝗻𝗮 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀! A new set of high-performance sentence embedding models, boasting between 35m to 6b parameters, expertly trained by @JinaAI_ and is accessible on 🤗. Excellent for neural search, reranking, and recsys 😎 arxiv.org/abs/2307.11224

📣 Announcing 𝐏𝐫𝐨𝐦𝐩𝐭𝐏𝐞𝐫𝐟𝐞𝐜𝐭 - a prompt-engineer-centric tool that automatically optimizes your prompts for ChatGPT, GPT3.5+, DALLE, Stable Diffusion! Prompt engineering done right! #LLMs #LMOps #LMs promptperfect.jina.ai

😬Poor Stanford undergrads, they would never thought they have inadvertently contributed to escalating tensions in the US-China AI race, geopolitical conflicts, and potential sanctions.

true. at least today most of chinese ai top mission is still about “look at me, i’m better then openai/ 🇺🇸 ” they don’t worry about the business model too much, that’s why they r from labs from giants alibaba, tencent, bytedance or this quant company.

hmm, this sounds too ambitious to me; but here is an interesting research proposal from @astar_research on Super Tiny Language Models (STLM) with only 10M, 50M, and 100M parameters. They aim to achieve competitive performance compared to models in the size range of 3B-7B parameters on GSM8K, MMLU, and LMSYS Chatbot Arena. Their starting point is a base model is a 10-layer llama2 (which gives random guess performance after training). They plan to explore more in two directions: 1. Model-level: weight tying, byte-level tokenization with a pooling mechanism, mixture of depths, layerskip, and text thought prediction. 2. Data-level: high-quality data selection and knowledge distillation. @astar_research Maybe also explore grokking here? arxiv.org/abs/2201.02177 Full paper here: arxiv.org/abs/2405.14159

Do you recognize this dataset? It is 𝗙𝗮𝘀𝗵𝗶𝗼𝗻-𝗠𝗡𝗜𝗦𝗧 created by me in 2017 when I was at @Zalando. Today, it just passed 10,000 @github stars and ~4,000 citations on @Google scholar!🎉 It has become one of the standard benchmark datasets in machine learning... 🧵

Here is an untrained Phi-3-tiny with 50M parameters inspired by Super Tiny LM paper from @astar_research @LeonGuertler. In their original proposal, Llama2 was the base so the tokenizer & activation function r different. Feel free to experiment it colab.research.google.com/dr…

How did we beat OpenAI's text-embedding-ada002 on 8K token length? When and why 8K token length matters to embeddings? Read our paper released today arxiv.org/abs/2310.19923

Exciting news! PromptPerfect now offers auto prompt engineering for #GPT4 and @LexicaArt - I'm really impressed by GPT4 for its complex reasoning and math problem-solving! Let's see some examples 🧵

here's a 2/3-step search example on "what is the latest blog post from jina ai", first reason to find jina ai news website, read its content, and determine the latest post

OpenAI's ChatGPT exclusive "Breeze" voice is impressively lifelike, better than any voice option in their public API.

People think LLMs hallucinate because of the knowledge cut-off, i.e., you are asking for new information created since training. This is true, but there is also another reason: you are asking for niche knowledge that has been "marginalized" during training. For example, when I asked GPT-3.5-turbo, "When was Jina AI founded?" I received an incorrect answer. This knowledge was definitely available before the cut-off date, but probably because we aren't that famous, it gets "marginalized" during training. This factuality issue can be easily solved with Jina Reader for search grounding, i.e., search the web, get the top 5 results, add them into the context window, and feed them to the LLMs. Seems reasonable. But how do those new clues really contribute to the correct answer? One could say the LLM is doing some in-context learning (ICL) on those new contexts, or does the LLM "wake up" the old memory and "fix" the "flawed distribution" because of this new context? This resonates with our recent work with @florianhoenicke on synthetic data generation for auto fine-tuning embedding models. Many people think this project is too good to be true: how can you train a good model but keep bootstrapping from a flawed distribution? How can you even generate hard-negatives? On this, I fully agree with @swyx's comment on synthetic data in one of @latentspacepod episodes: "The goal of synthetic data is less to emulate human speech; it is more to spike the distribution in useful ways." In the context of auto fine-tuning embedding models, synthetic hard-negatives do not have to be perfect; they just have to be a "wake-up call" for the model and steer it in the direction that the user points to.

I've been diving into DSPy last few days, and although it's impressively powerful, I've hit a few bumps trying to wrap my head around some concepts. Here's my take on it: 1⃣ What are teleprompter, optimization, and compile? What's exactly being optimized? In DSPy, "Teleprompters" is the optimizer, (and looks like @lateinteraction is revamping the docs and code to clarify this). The `compile` function acts at the heart of this optimizer, akin to calling optimizer.optimize(). Think of it as the DSPy equivalent of training. This process aims to tune (1) the LM weights, (2) the instructions, and (3) few-shot demonstrations. However, most beginner DSPy tutorials won't delve into adjusting (1) & (2), leading to my next query: 2⃣ What's bootstrap all about? Bootstrap refers to the creation of self-generated demonstrations for few-shot in-context learning, a crucial part of the compile (i.e., optimization/training) process. These few-shot demos are generated from user-given labeled data; and one demo often consists of input, output, rationale (e.g., in Chains of Thought), and intermediate inputs & outputs (for multi-stage prompts). Of course, quality few-shot demos are key to the output excellence. To that, DSPy allows user-defined metric functions to ensure only demos that meet certain criteria are chosen. The metric function and its behavior, well, needs more than another tweet to talk about.

Do you like our new @github organization page? 🤩Check it out: github.com/jina-ai #opensource #neuralsearch

What's a better way to spend Friday night to support FastAPI and pydantic? Many! But no regret.😂 DocArray 0.1.8 now supports data validation and now you can seamlessly use it in FastAPI to build reliable webservices! docarray.jina.ai/fundamental… @tiangolo @samuelcolvin

> Temperature: 0.6 i mean we r already here, so why not just 0.618 add some 🧙 to it

Sorry to be the buzz killer this #AutoGPT party. Here is my unpopular opinion about it. Today, I had a time to look at its source code and play it with my colleagues at @JinaAI_ , here is what I learned 👇 jina.ai/news/auto-gpt-unmask…

They kinda apply this guilty-until-proven-innocent logic to OSS, which is understandable in US but indeed very uncomfortable

need to delve into it.

here's a 13-step query on "who is the biggest, cohere, jina ai, voyage" - after search, reflect, looping, the result is correct. it's cohere. video in 2x speedup

🧁PromptPerfect 0.10 has just been released! With Kandinsky 2.1 support, a new text2image model from @sberbank AI lab; and the new template management system! Try it now, and be impressed by Kandinsky's image quality and speed! promptperfect.jina.ai

the biggest problem of german society is obsession with mediocrity - hard working and being smart r treated like sins - so i don’t understand why AFD will make Germany any better, just another bunch of average people.

While migrating from davinci003 to ChatGPT API @OpenAI released yesterday, we found two interesting observations. Good or bad? u tell me. First, the `assistant` role in the new API always addresses itself in the first person. This can be convenient in conversation UX, but in

🙌Join 07/23 🇦🇹 Vienna. 𝐂𝐫𝐞𝐚𝐭𝐢𝐯𝐞 𝐀𝐈 & 𝐀𝐫𝐭 is the perfect event for AI engineers, artists and enthusiasts! This #Saturday, don't miss out on the chance to explore the intersection of creativity and artificial intelligence. #CreativeAI #Art meetup.com/jina-community-me…

Okay I hear u: we go 3D! So let'em rebuild my fav 3D maze screensaver from win95! Deepseek-r1, o3-mini-high, claude-3.5-sonnet, which is the best? Tougher than my last letter dropping animation as 3D stuff + maze generation + maze solving! All results are one-shot. Full prompt: Create a complete, self-contained HTML file for an auto-navigating 3D maze screensaver that: 1. Structure: - Put all HTML, CSS, JavaScript and Three.js in one file - Import Three.js from Cloudflare CDN - No external assets or dependencies besides Three.js 2. Core Features: - Simple first-person camera that auto-walks through red brick corridors - Tan floor, white ceiling (hex colors 8B4513 for floor, FFFFFF for ceiling) - Fixed 90-degree turns using right-wall following - Basic lighting and fog for depth - Screen-filling canvas (100vw/100vh) 3. Technical Requirements: - Generate a 10x10 maze using a simple array-based structure - Smooth camera transitions (lerp) for walking and turning - Regenerate new maze when solved - 60fps target performance - No user controls needed, pure auto-navigation Please provide a complete code that runs immediately when the page loads, with all maze logic self-contained. The file should work by simply opening it in a browser.

Just attended @seb_ruder's amazing tutorial on parameter-efficient fine-tuning at #EMNLP2022 and my mind is blown! His guidance has given me so many ideas for the next release of Finetuner. Thank you for sharing your wisdom, Seb!

In their latest demo, @OpenAI unveiled the impressive multimodal capabilities of #GPT4, generating text descriptions from images with ease. Give PromptPerfect 0.6 a spin to experience this feature firsthand! Spoiler: so much better than #BLIP2! Let's see some examples, 🚀🧵

Friday night #DiscoArt with github.com/jina-ai/discoart 🧵🎨 DocArray ID attached: 'discoart-1490247652'

huh? ChatGPT Plus is intrinsic `text-davinci-002`??? Thought it should -004 or something, not even davinci-003? @OpenAI

black hole simulator by qwen3-coder, unsure about the physics behind but looks pretty cool

3 months ago, I announced my new venture on Neural Search. Today, I’m proudly revealing 🔍Jina: the cloud-native neural search framework powered by state-of-the-art AI & deep learning. Think out-of-the-[text]box, use Jina to build your next search system! github.com/jina-ai/jina

“找不到对象”

What should we learn from ModernBERT?

First post of 2019🎊: serving Google BERT model in production using Tensorflow and ZeroMQ, where I explain the design philosophy behind my open source project "bert-as-service". If your new year’s resolution is putting your ML/AI p…lnkd.in/fNpGDvK lnkd.in/fPTFjVu

LLM/ChatGPT hype today in one pic. If one really wants to build a decentralized LLM, 𝐭𝐡𝐞 𝐨𝐰𝐧𝐞𝐫𝐬𝐡𝐢𝐩 𝐨𝐟 𝐭𝐡𝐢𝐬 𝐋𝐋𝐌 𝐬𝐡𝐨𝐮𝐥𝐝 𝐛𝐞 𝐝𝐢𝐬𝐭𝐫𝐢𝐛𝐮𝐭𝐞𝐝 𝐭𝐨 𝐞𝐯𝐞𝐫𝐲 𝐩𝐞𝐫𝐬𝐨𝐧 𝐰𝐡𝐨 𝐜𝐨𝐧𝐭𝐫𝐢𝐛𝐮𝐭𝐞𝐬 𝐭𝐨 𝐭𝐡𝐞 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐜𝐨𝐫𝐩𝐮𝐬, even if they write one sentence.

Now we are talking

the graph illustration looks pretty awesome ngl

It’s true, but the idea of "training on YouTube" is a major misconception in the West. Here’s why: 1. Most Chinese users aren’t thrilled with Westernized GenAI image and video results because they lack cultural roots and elements. For example, they might wonder why a guy has blond hair or why a temple doesn’t look like theirs. This "foreign-feeling" model quickly loses appeal among Chinese GenZ. 2. China’s "intranet" has huge content platforms too, like Kuaishou for short videos and Bilibili for long videos. Content on these platforms is fully moderated and censored, making it much safer for Chinese AI companies to train on domestic data. 3. There’s also top-down AI/model/IP regulation in China. It’s real, but drawing a clear line is tough because politics play a big part.

So so much fun to practice Jina (github.com/jina-ai/jina) in a GPU-heavy pipeline consisting of #DALLE-mega @borisdayma, GLID, SwinIR, and CLIP-as-service. Can't stop 😁 Early next week I will share a super-easy notebook for you to reproduce. 🧵

🚀Unleash the power of visual storytelling with #SceneXplain 0.2! Introducing the new & fast "Comet" algorithm to solve hallucination issues & elevate your image narratives. Now with batch support for 128+ images in a single shot! Experience it today: scenex.jina.ai

We benchmarked @midjourney /describe command released earlier today vs. SceneXplain released yesterday vs. CLIPInterogator 2.1 and BLIP2 on image captioning and "reverse-engineering" prompts, here is what we learned. Full article jina.ai/news/scenexplain-unl…

Looking at our DeepSearch code, I found rightly anticipated needs, unnecessary needs wrongly expected, and new needs we hadn't foreseen: 🐐 GOAT: Long-context LLM, JSONSchema following, query expansion, query dedup with je-v3 (this unexpected), websearch & read (jina reader), @vercel ai sdk 😐 Nah: reasoning LLM, rerankers, SLM for query expansion 🗑️ Not used: query routing, vectordb, agent framework, agent-memory framework

Super excited to announce DALL·E Flow: a Human-in-the-Loop workflow for creating HD images from text. Special thanks to @borisdayma @rom1504 for their support over the weekend! #opensource #dalle Now, look at the amazing results that you can get in 🧵 github.com/jina-ai/dalle-flo…

.@kevinsxu found that when running Qwen/Deepseek locally, these models are not "censored" and are quite capable of answering questions about Tiananmen or commenting on Xi. The guardrails only appear as post-hoc rules on their cloud API. So those eastern models “know” just as much as their western counterparts. interconnected.blog/was-zuck…

为什么不直接prompting给GPT？因为这里用到了Langchain实现的ReAct framework实现Agent之间推理和协同。我司上个月也有类似的项目github.com/jina-ai/agentchai…

i think they “apologized”? aksh-garg.medium.com/llama-3…

Jina: github.com/jina-ai/jina because I’m the creator of it.

这里的观点和我之前写的一篇文章一样，ChatGPT没有帮助SEO而是直接杀掉了SEO，当然取而代之的是LLMO，或者学术圈叫做in-context learning jina.ai/news/seo-is-dead-lon…

This one credits to @Alibaba_Qwen @JustinLin610 team for providing the best base model we could have, Qwen2.5-1.5B-Instruction, very strong long-context performance.

"Bootstrapped 0 full traces after 20 examples in round 0" is probably the most frustrating message for DSPy newbies. This silent error essentially means that optimization/compilation failed ❌, and the prompt you get is no better than simple few-shot. What goes wrong ? I've summarized some tips to help you debug ur DSPy program when encounter such message: 🧵

It is hard to tell if people hate to love RAG or love to hate RAG. According to recent discussions on X and HN, RAG should be dead again. This time, critics are focusing on the over-engineering of most RAG frameworks, which, as @jeremyphoward @HamelHusain @Yampeleg demonstrated, could be accomplished with 20 lines of Python code. The last time we had this vibe was shortly after the release of Claude/Gemini with a super long context window. What makes this time worse is that even Google's RAG generates funny results as @icreatelife @mark_riedl showed, which is ironic because back in April, at Google Next in Las Vegas, Google presented RAG as the grounding solution. I see two problems with the RAG frameworks and solutions we have today. First, nearly all RAG frameworks implement only a "feed-forward" path and lack a "back-propagation" path. It is an incomplete system. I remember @swyx, in one of the episodes of @latentspacepod, arguing that RAG will not be killed by the long context window of LLMs since (1) long context is expensive for devs and (2) long context is hard to debug and lacks decomposability. But if all RAG frameworks focus only on the forwarding path, how is it easier to debug than an LLM? It is also interesting how many people get overexcited by the results of RAG in some POCs and forget that adding more forward layers (in RAG's case, the LLM generator) without backward tuning is a terrible choice. We all know that adding one more layer to your neural networks expands its parametric space and hence representation ability, enabling it to do more potential things, but without training, this is nothing. There are quite some startups in the Bay Area working on evaluation—essentially trying to evaluate the loss of a feed-forward system. Is it useful? Yes. But does it help close the loop of RAG? No. So who is working on the back-propagation of RAG? Afaik not many. I am mostly familiar with DSPy, a library from @stanfordnlp @lateinteraction that sets its mission on that. But even for DSPy, the main focus (or community usage) is on optimizing few-shot demonstrations, not the full system. So why is this problem difficult? Because the signal is very sparse, and optimizing a non-differentiable pipeline system is essentially a combinatorial problem—in other words, extremely hard. I learned some submodular optimization during my PhD, and I have a feeling that this technique will be put to good use in RAG optimization. Second, I do agree that RAG is for grounding, despite the funny search results from Google. There are two types of grounding: search grounding, which uses search engines to extend the world knowledge of LLMs, and check grounding, which uses private knowledge (proprietary data) to do fact-checking. In both cases, it cites external knowledge to improve the factuality of the result, provided that these external resources are trustworthy. In Google's funny search result, one can easily see that not everything on the web is trustworthy (yeah, big surprise, who would thought!), which makes search grounding look bad. But I do believe you can only laugh at it for now. There are some implicit feedback mechanisms behind the Google Search UI that collect users' reactions to those results and weight the credibility of the website for better grounding. In general, it should be pretty temporary, as this RAG just needs to get past the cold start, and results will improve over time. My take is that RAG is neither dead nor alive; it is just one algorithm pattern you can use. If you make it the algorithm and idolize it, then you are living in a bubble you created, and the bubble will burst.

clippy: that eyes are mine and give back my eyebrows!

😂 one day AI will reply our stupid test prompt like “don’t u have a work to do? ahh i forgot i took it, sorry”

直接用PromptPerfect promptperfect.jina.ai 还可以对image2prompt逆向工程

🚀PromptPerfect 0.9 now supports auto prompt engineering for Claude from @AnthropicAI. Imo, Claude is a smart AI but often too "stuffy": too polite, serious, can't joke. Now PP loosened it up 😁Try it yourself. promptperfect.jina.ai

man, gemini-flash-2.0 search grounding is no joke. I cherry-picked some very niche questions about jina ai that only I know the answers to and used them for evaluation. I uploaded them to GitHub yesterday for the community to benchmark, and today Gemini has already found them and used them as shortcut and achieved 100% accuracy🤦‍♂️ So never publish your benchmark to internet in plain text.

🚀Blast off with PromptPerfect 0.20! New onboarding is as easy as pie 🥧, more tutorial videos than you can shake a stick at! This isn't just an update; it's a whole level-up! promptperfect.jina.ai Caution: side effects include exceptional PE skills🛠️ and irresistible charm. 😎

In Jina 2.0 (github.com/jina-ai/jina), we extend #FastAPI #SwaggerUI to pretty-print multimedia responses into listview & flowchart. This allows Jina developers to play and test REST endpoints in prompt. Try it now!

Sat.night 2 hours on porting @GoogleAI latest BiT model in (Poké-)production using @JinaAI_. Now you can look for similar @Pokemon with SOTA visual representation learning. Feat. replicas and shards, containerization, REST & gRPC gateway, you name it! github.com/jina-ai/examples/…

SEO is dead as #ChatGPT is replacing Google search, so how can businesses ensure their info appears in its single answer? In-context learning on LLMs is the solution. jina.ai/news/seo-is-dead-lon… Special thanks to @GaryMarcus @seb_ruder @Nils_Reimers for inspiration at @emnlpmeeting

Introducing 𝗥𝗮𝘁𝗶𝗼𝗻𝗮𝗹𝗲: a decision-making tool powered by GPT3.x and in-context learning for analyzing Pros & Cons, SWOT. Perfect for managers, business owners, and individuals. 🌍Multilingual support! 🌟Try it now rationale.jina.ai #GPT3 #GenerativeAI

My hot takes on DSPy, why it's crucial for future prompt engineering, and yet why it's challenging for average prompt engineers to learn.

📷Discover the power of #GPT4 -like multimodal image explaining with SceneXplain! Advanced image storytelling driven by LLMs, tailored for complex scenes & multilingual support. Fast batch processing with our API. Don't wait; elevate your visuals today! scenex.jina.ai

PromptPerfect 0.4 now improves the auto prompt engineering and supports @midjourney You can now easily generate stunning images 𝙬𝙞𝙩𝙝𝙤𝙪𝙩 carefully designed prompts.

A big & solid 1.0 after a year of hard work! Congrats to our team and OSS community! ❤️ @PythonWeekly @ThePSF @PyData @pydataberlin @pycon @gvanrossum @pycoders @PythonHub @pypi @kdnuggets And a big thank you to our upstream: @numpy_team #Protobuf #gRPC @libzmq #PyYAML #FastAPI

Thanks to @willmcgugan #Rich, DocArray nesting and embedding visualization just got way cooler. github.com/jina-ai/docarray

comparing to davinci003, ChatGPT API implementation of ethical filtering can be excessive, as it tends to include disclaimers and even censor friendly sarcasm. Should a competitor arise with a less restrictive alternative, this could be the downfall of ChatGPT API.

也可以把复制出来的prompt放到promptperfect.jina.ai 作进一步的优化

🎉Introducing my latest work: GNES!🍾GNES is Generic Neural Elastic Search, a cloud-native semantic search system based on DNN, it enables large-scale index and search for text-to-text🔠, image-to-image🖼️, video-to-video🎞️and any-to-any content form. github.com/gnes-ai/gnes

BERT-as-service is honored to be one of the most popular open source project of the past year. Thanks all for making it happen!🤜🏻🤛🏻🎊

Buying a Porsche 911🏎️ or a house in Spain🏡? Let 𝐑𝐚𝐭𝐢𝐨𝐧𝐚𝐥𝐞 0.2 choose the best for you. The new multi-option analysis uses in-context learning and latest #GPT3.x to generate a multi-criteria analysis that ends your indecisiveness rationale.jina.ai #GenerativeAI

Today, I visited a big tech in Bay, a key user of jina-embeddings-v2. Their feedback was encouraging: "We don't rely on public MTEB; we assess all embeddings on our domain data, and Jina is the best."🥹 At @JinaAI_, we focus on solving real problems, not just meeting benchmarks.

只是模仿英文语音的话，推荐Coqui AI，柏林隔壁一家初创，创始人之前做Firefox TTS的。目前图形化界面和开源的都有。我司之前拿这个模仿过Morgan Freeman，效果你们说

2024 how to prevent hallucinations 2025 how to prevent overthinking

🔥 SceneX vs. miniGPT4: New benchmark results reveal SceneX's superior performance in image captioning! 🚀 SceneX consistently outperforms miniGPT-4 in capturing intricate visual details and generating engaging captions. Full breakdown in this thread!👇jina.ai/news/scenexplain-vs-…

还是自然语言，直接走一层pp即可，不仅gpt, dalle, midjourney 都可以用 promptperfect.jina.ai

Thanks @OpenAI. This time is quick ⏩⏩

If you still think cosine similarity between document embedding and query embedding for search, think again! Dive into ColBERT now:

Back from @KubeCon_ and unpacking, I don’t think I need any socks 🧦 this summer. Thanks a lot for these creative and colorful swags 🌈 @DataStax @runailabs @elastic @InfluxDB @ocrasec @aiven_io @zesty_co @era #KubeCon #CloudNativeCon #KubeConEU

automatic prompt engineering promptperfect.jina.ai

this is ur Deep Research guy