Tweets

BrowseSafe is a specialized detection model to defend against evolving prompt injection attacks.

It is designed specifically to spot and block malicious instructions hidden in web pages before they can impact AI browser agents.

huggingface.co/perplexity-ai/…

Caption this! What’s the perfect name for this photo?

Large-scale infrastructure buildout is essential to the next decade of American Dynamism. Making it happen will require leaps in capabilities; construction projects today are often slower and more expensive than they were 50 years ago. America’s spirit of speed, precision, and ambition has been swallowed by incremental, risk-averse, and expensive processes.

Unlimited Industries is a vertically integrated, AI-native platform that designs and builds critical infrastructure like power plants, factories, and data centers from end-to-end. Unlike the status quo of misaligned cost-plus contracts, Unlimited aligns incentives to results and compresses design cycles to unlock dramatic increases in speed and efficiency.

We’re honored to co-lead @unltdindustries’s $12M Seed round and partner with @alexmodon, @_jordanstern, and @taraviswanathan. Let’s build.

I'd like to share some insights on using AI to translate articles. Here is the core philosophy:

The best translation is actually rewriting.
Achieving high-quality results requires breaking the process down into steps.

However, this depends on the scenario. For general translation needs, a single rewrite is usually sufficient. Given the capabilities of current Large Language Models—especially models like Gemini 3 Pro—the quality of a one-shot rewrite is already quite impressive.

If you are aiming for professional-grade translation, it is necessary to have the AI proofread and polish the text after that initial rewrite.

Crucially, do not attempt to cram translation, proofreading, and polishing into a single prompt, unless the content is extremely short.

The main reason is something I mentioned yesterday: while models can handle very long inputs, they tend to cut corners and suffer from severe hallucinations if the required output is too long.

Imagine translating a 2,000-word article. If you ask the model to rewrite, then proofread, and finally polish it all at once, the output could easily balloon to 5,000 - 6,000 words. By the time it gets to the end, the quality will degrade significantly.

Therefore, translation, proofreading, and polishing are best performed independently.

Let's start with translation. The prompt doesn't need to be overly complex. Simply asking it to "rewrite" works well, provided you specify the following:
- The style of the article
- A glossary of common terms or mappings
- The target audience (optional)

For proofreading, you need to provide both the original text and the translation. The goal here is to check for omissions and errors. If your requirement for precision isn't extremely high, you can skip this step.

For polishing, you do not need the original English text anymore; the translation alone is sufficient. At this stage, you just need the model to check if the results flow smoothly and if the phrasing adheres to native expression habits. The original text is no longer necessary.

Additionally, if the article is too long, you should split it into blocks. It is best to split by natural chapters or paragraphs; usually, splitting at the paragraph level is enough.

As for how to ensure continuity between the current block and the previous one, a simple and effective method is to add the source text and translation of the previous block into the context. This allows the translation of the next block to reference the content and style of the one before it.

Regarding how much history to retain, that depends on the model and the size of the blocks. Usually, the source and translation of the immediately preceding block are enough; you don't need too much history. In fact, sometimes providing no history at all works fine.

This process can be automated via a program using an API, or done manually within the model interface. Personally, I use Gemini. I create different "Gems" for different prompts, and when I need them, I simply paste the content into the corresponding Gem.

Gemini knows your location and current date, so you can ask gemini to get the location and date by itself, e.g.
----
City name: {get my location from my profile}
Date: {get current date}
----

------ full prompt -------

Present a clear, 45° top-down view of a vertical (9:16) isometric miniature 3D cartoon scene, highlighting iconic landmarks centered in the composition to showcase precise and delicate modeling.

The scene features soft, refined textures with realistic PBR materials and gentle, lifelike lighting and shadow effects. Weather elements are creatively integrated into the urban architecture, establishing a dynamic interaction between the city's landscape and atmospheric conditions, creating an immersive weather ambiance.

Use a clean, unified composition with minimalistic aesthetics and a soft, solid-colored background that highlights the main content. The overall visual style is fresh and soothing.

Display a prominent weather icon at the top-center, with the date (x-small text) and temperature range (medium text) beneath it. The city name (large text) is positioned directly above the weather icon. The weather information has no background and can subtly overlap with the buildings.

The text should match the input city's native language.
Please retrieve current weather conditions for the specified city before rendering.

City name: {get my location from my profile}
Date: {get current date}

Prompt for High-Quality Translation

---- Prompt ----

Please rewrite the following  article into accessible, fluent, and engaging {{Simplified Chinese}}.

Core Requirements:
- Audience & Style: Target general readers interested in AI. Adopt a storytelling style that is clear and easy to understand, rather than the tone of an academic paper.

- Accuracy First: Core facts, data, and logic must remain strictly consistent with the original text.

- Flow & Fluency: Prioritize authentic target language sentence structure. Break down long, complex sentences into more natural, shorter phrases.

- Standard Terminology: Use industry-recognized standard translations for technical terms (e.g., {{overfitting -> 过拟合}}). Upon the first occurrence of a term, include the original text in parentheses.

- Preserve Formatting: Maintain the original Markdown formatting, including headings, bolding, italics, and images.

- Contextual Annotation: If technical terms are difficult for a layperson to understand, or if missing background knowledge hinders comprehension, add appropriate annotations to facilitate understanding. Wrap these annotations in bold parentheses like this: (annotation).

Glossary / Common Vocabulary:
{{- AI Agent -> AI 智能体
- LLM -> 大语言模型}}

Now, please rewrite the content below:

See how frontier MoEs like DeepSeek-R1 scale with extreme co-design. With NVIDIA Blackwell and NVIDIA Dynamo, enterprises get 10× performance and revenue efficiency at lower cost per token. And now, with Dynamo integrated across @awscloud, @Azure, @googlecloud, and @OracleCloud, that same performance is available to everyone. 

Learn more: nvda.ws/4hUXcki

In the latest issue of Research Focus: Fara-7B brings efficient agentic computer use to small models; new tools help computing researchers assess societal impact; and more. msft.it/6017tbeFf

nano banana pro prompt

Transform a comic character into an ultra-realistic human

---- prompt ---- 
Transform this comic character into an ultra-realistic human while preserving the original hairstyle, outfit, facial expression, and overall character identity.
The entire scene should use deep depth of field, keeping both the model and the environment extremely sharp, creating an immersive, cinematic smartphone photography look.

STYLE:
- Cinematic ultra-realistic fashion photography
- High-resolution smartphone camera aesthetic with crisp, sharp details
- Dramatic lighting contrast between warm work lights and cool twilight tones
- the model is the main subject while retaining rich environmental details

TECHNICAL SPECS:
- Camera: flagship smartphone camera
- Lens: standard built-in phone lens
- Aperture: f/8–f/11 for deep depth of field
- Resolution: 4K or higher

NEGATIVE PROMPT:
- blurry background, shallow depth of field, bokeh
- out of focus, distorted face
- cartoon, anime, CGI character, illustration, painting look
- low quality, pixelation, noise
- harsh direct sunlight or overexposed lighting

𝐌𝐢𝐠𝐫𝐚𝐭𝐢𝐧𝐠 𝐟𝐫𝐨𝐦 𝐄𝐥𝐚𝐬𝐭𝐢𝐜𝐬𝐞𝐚𝐫𝐜𝐡 𝐭𝐨 𝐐𝐝𝐫𝐚𝐧𝐭 - 𝐃𝐞𝐞𝐩-𝐃𝐢𝐯𝐞 𝐛𝐲 Mahimai Raja J

We think this is a a great technical breakdown on why many teams are moving from Elasticsearch to Qdrant for their vector search workloads.

Why he wrote this?

After facing challenges with scaling Elasticsearch for vector workloads - complex configs, higher infra cost, and limited vector performance - Mahimai created a practical guide to help teams transition smoothly to a vector-first stack.

𝐈𝐭 𝐜𝐨𝐧𝐭𝐚𝐢𝐧𝐬:

𝑾𝒉𝒚 𝒎𝒊𝒈𝒓𝒂𝒕𝒆?
- Challenging to build scaled, performant vector search with Elastic
- Need to reduce latency
- Increase resource efficiency

𝑯𝒐𝒘 𝑸𝒅𝒓𝒂𝒏𝒕 𝒉𝒆𝒍𝒑𝒔?
- Native vector indexing
- Strong payload filtering
- Efficient dense + sparse hybrid search
- Easier scaling and maintenance

𝑻𝒉𝒆 𝒎𝒊𝒈𝒓𝒂𝒕𝒊𝒐𝒏 𝒑𝒓𝒐𝒄𝒆𝒔𝒔:
- Exporting ES data & embeddings
- Re-mapping schema for Qdrant
- Rebuilding collections & payloads
- Updating query patterns for vector search
- Handling ranking and scoring differences

Practical guidance:

Includes repub.towardsai.net/how-to-migrate… snippets, and common pitfalls to avoid during migration.

Full article 👉 https://t.co/6FVNaQzFbO

#Qdrant #VectorSearch #Migration #SearchEngineering #LLMApps