Archive

Nano Banana — Google's Next-Gen Image Editing AI (Free to Try) | Nano Banana

Precise, controllable, and consistent image edits with mask-free inpainting, layout-aware outpainting, and conversational refinement.

See examples on the page.

Farer

1. Charlton Green - 38mm Swiss-Made GMT Watch With Sapphire Bezel
2. Mansfield Cushion Case - 38.5mm Swiss-Made Manual Wind Watch
3. World Timer

llama.cpp guide: running gpt-oss with llama.cpp

brew install llama.cpp
llama-server -hf ggml-org/gpt-oss-20b-GGUF \
  --ctx-size 0 --jinja -ub 2048 -b 2048 -ngl 99 -fa

This downloads a 12GB model file from ggml-org/gpt-oss-20b-GGUF on Hugging Face, stores it in ~/Library/Caches/llama.cpp/ and starts it running on port 8080.

You can then visit this URL to start interacting with the model: http://localhost:8080/

shortcut.ai -- AI powered Excel sheet. Not free $40/mo min.

Came to know about Task specific pruning via this X/MaximeRivest post. See (Reda et al. 2025)

Pruning might be useful for niche languages. (I am thinking of Nim after reading this thread earlier)

References

• Reda, Jangda, Chintalapudi (2025). Task Specific Pruning with LLM-sieve: How Many Parameters Does Your Task Really Need?.

Using OpenAI API for Sentiment Analysis by Ashwin Malshe; Code - GitHub:ashgreat/Sentiment-Analysis-OpenAI. The usual/classic way of doing Sentiment Analysis is using NLP.

#todo #exercise rewrite the same using dspy.

Kindle Modding Wiki

jailbreaking it takes 5 minutes and turns it into one of the cheapest, most capable eInk devices you can own. Runs Linux! -- via x/dqmonn

I have been using Hetzner for a while, and this page is to keep notes on things that might be beneficial on Hetzner.

Why use Dokploy -- DIY Serverless (with coreOS + dokploy)

load that VPS with a bomb-proof no-config OS like coreOS, and put Dokploy on it for a no-fuss container service UI. It gives you the hassle-free experience of a Heroku-like platform, but with the cost and independence of a VPS.

This way you can have your cake and eat it too: most of the benefits of serverless, with most of the benefits of running your own VPS.

Similar software mentioned in this HN discussion of the above article: Harbormaster

Via Jason Liu on LinkedIn

The gap between AI demos and production isn't about reasoning quality anymore, it's about context.

Three critical lessons worth stealing:

1. Domain experts are your secret weapon

• Build custom UIs that optimize for high-quality reviews, not just volume • Collect four outputs from every review: performance metrics, failure modes, suggested improvements, and input-output pairs • Use these to create failure mode datasets you can continuously test against

2. Prompting beats fine-tuning for vertical AI

• Modern LLMs already have strong domain reasoning, they lack customer-specific context • Skip the brittle prompt engineering and focus on context augmentation instead • Use domain knowledge retrieval at inference time to inject customer-specific definitions and workflows

3. Trust requires more than good performance

• Monitor production outputs with domain expert reviews (start with humans, not LLM judges) • Define clear sampling strategies as you scale: uncertainty-based, outlier detection, stratified sampling • Have response protocols ready when performance dips below SLAs

An important question to ask is: What's your system for incorporating domain expertise?

Teams pay too much attention to model selection but neglect the process for continuously baking in expert knowledge. Reliable vertical AI implementations aren't just about picking the right model, they're about building flywheels that systematically improve with use.

LanceDB · GitHub where

1. lancedb/lance is "Modern columnar data format for ML and LLMs implemented in Rust. Convert from parquet in 2 lines of code for 100x faster random access, vector index, and data versioning. Compatible with Pandas, DuckDB, Polars, Pyarrow, and PyTorch"
2. lancedb/lancedb: is "embedded retrieval engine for multimodal AI."

lake house = data lake + data warehouse.

stores raw, unstructured, and semi-structured data cheaply, often on object-storage like a data lake.

supports ACID transactions, structured data, schema enforcement, and fast SQL based Analytical Queries.

Basically, lakehouse = store all your data, and run different kind of loads on them - analytics, machine learning, Bi etc, without moving around data.

Common solutions include: Databricks lakehouse, Apache Iceberg, Apache Hudi, Delta Lake, LanceDB etc.,

Qwen-Image lightning generation on Apple Silicon:

$ uv run https://raw.githubusercontent.com/ivanfioravanti/qwen-image-mps/refs/heads/main/qwen-image-mps.py -f -p "Generate an SVG of a pelican riding a bicycle"

This will download 1.7GB of Lora in the folder you start it from. 9:35:30-

--

Qwen-Image-Edit: Image Editing with Higher Quality and Efficiency; Aug 2025.

uv run https://raw.githubusercontent.com/ivanfioravanti/qwen-image-mps/refs/heads/main/qwen-image-mps.py generate -f -p "A rollercoaster with view on Carmel, IN Midtown"

(Vaibhav Gupta and team 2024) is a simple prompting language for building reliable AI workflows and agents by BoundaryML. GitHub - BoundaryML/baml: The AI framework that adds the engineering to prompt engineering. Available for Python, typescript, golang, java etc.,

BAML makes prompt engineering easy by turning it into schema engineering -- where you mostly focus on the models of your prompt -- to get more reliable outputs.

New syntax can be incredible at expressing new ideas. Plus the idea of maintaining hundreds of f-strings for prompts kind of disgusts us 🤮. Strings are bad for maintainable codebases. We prefer structured strings.

The goal of BAML is to give you the expressiveness of English, but the structure of code.

Half the error rate, just by changing the way the schema is expressed to the LLM ! I think the lesson learned here is that:

1. Easier to read for humans usually means easier to read for LLMs as well
2. JSON is not easy to read for humans -- X/appenz

The new BAML adapter (which subclasses the existing JSON adapter) avoids using JSON schema for formatting the data model in the prompt. Instead, it uses the format popularized by BAML, and it instantly improves the quality of structured outputs. -- X/tech_optimist Using the BAMLAdapter in dspy is incredibly simple - just pass it as a parameter to your LM, and the adapter takes care of the rest.

import dspy
from pydantic import BaseModel, Field
from typing import Literal
from dspy.adapters.baml_adapter import BAMLAdapter

# 1. Define Pydantic models
class PatientAddress(BaseModel):
    street: str
    city: str
    country: Literal["US", "CA"]

class PatientDetails(BaseModel):
    name: str = Field(description="Full name of the patient.")
    age: int
    address: PatientAddress | None

# 2. Define a signature using the Pydantic model as an output field
class ExtractPatientInfo(dspy.Signature):
    '''Extract patient information from the clinical note.'''
    clinical_note: str = dspy.InputField()
    patient_info: PatientDetails = dspy.OutputField()

# 3. Configure dspy to use the custom adapter
llm = dspy.OpenAI(model="gpt-4.1-mini")
dspy.configure(lm=llm, adapter=BAMLAdapter())

# 4. Run your program
extractor = dspy.Predict(ExtractPatientInfo)
note = "John Doe, 45 years old, lives at 123 Main St, Anytown. Resident of the US."
result = extractor(clinical_note=note)
print(result.patient_info)

# Expected output:
# PatientDetails(
#    name='John Doe',
#    age=45,
#    address=PatientAddress(street='123 Main St', city='Anytown', country='US'))

References

• Vaibhav Gupta, team (2024). BAML.

Collection of "Awesome X" collections.

• Shubhamsaboo/awesome-llm-apps: Collection of awesome LLM apps with AI Agents and RAG using OpenAI, Anthropic, Gemini and opensource models.

• Collection of LLM papers: GitHub - horseee/Awesome-Efficient-LLM: A curated list for Efficient Large Language Models by Xinyin Ma.

• Microsoft LMOps -- a research initiative on fundamental research and technology for building AI products w/ foundation models, especially on the general technology for enabling AI capabilities w/ LLMs and Generative AI models.

Extracting structured information from unstructured text using LLMs is a good tool to have in your llm toolkit.

Google's langextract is one such python library for extracting structured information from unstructured text using LLMs with precise source grounding and interactive visualization.

The other popular library is Instructor -- extracting structured data from Large Language Models (LLMs). Built on top of Pydantic, Instructor provides type-safe data extraction with automatic validation, retries, and streaming support. supports all major providers like openai, claude and local-llm like ollama etc.,

What if you do not want to use AI SaaS products for any reason (monetary - subscriptions add up quick, privacy - do not want share your data with companies, etc.,)?

One approach is to use local-llm on your laptop, or pc. You might also be ok with doing so on a cloud hosted instance where you get to control the entire stack.

Another approach is to use some of the "free tier" as seen here -- How I Code with AI on a budget/free; via HN.

Every month so many new models come out. My new fav is GLM-4.5... Kimi K2 is also good, and Qwen3-Coder 480b, or 2507 instruct.. very good as well. All of those work really well in any agentic environment/in agent tool

Extracting content, and retaining all the semantic information presented in texts, tables, charts, graphs, and other visual elements in a manner that is most suitable for consumption by LLMs, especially when used in conjunction with RAG is still a hard problem.

There are many compnanies, research teams and open source projects attempting to solve this at the document level and as a pipeline.

Some of the ones that are the top of my head right now (2025-08) are:

Open Source Projects

• Docling by IBM Zurich Research, and SmolDocling; see (Team 2024)
• Dolphin by ByteDance; see (Feng et al. 2025)
• roe-ai/vectorless by AI SaaS -- "PDF chatbot that uses no vector embeddings or traditional RAG. Instead, it leverages Large Language Models for intelligent document selection and page relevance detection, providing a completely stateless and privacy-first experience."

Supporting Libraries

• pypdf2
• GitHub - pypdfium2-team/pypdfium2: Python bindings to PDFium, reasonably cross-platform. (mentioned in docling paper, as well as used in datalab-to/marker)
• #todo (one used in tensorlake)

Companies

See Running Notes on RAG

• Datalab; see GitHub - datalab-to/marker: Convert PDF to markdown + JSON quickly with high accuracy

References

• Team (2024). Docling Technical Report.
• Feng, Wei, Fei, Shi, Han, Liao, Lu, Wu, Liu, Lin, others (2025). Dolphin: Document Image Parsing via Heterogeneous Anchor Prompting.

Streamlit is a python library to build data apps. This is similar to observable, and evidence.

With uv, it is easy to see what streamlit looks like -- uv run --with streamlit streamlit hello -- this launches a demo app with a couple different visualization screens.

This is more along the lines of "UI controls that allow you to change the parameters of a chart/visualization" than evidence's "Dynamic report" approach.

Streamlit is a pythonic middle ground between reporting oriented evidence and a very visualization oriented observable.

another way to look at these three: streamlit is python, observable is javascript, and evidence is sql.

Evidence is "Business Intelligence as Code"; which means it allows you to build reports, and standard business data visualization ( charts, and basic graphics) using sql with markdown as the glue between various data and visualization elements. You can put to gether a dynamic report with a few interactive elements (e.g: filter data, change date ranges) that can be exported as "static" web pages. The exported web pages will also have interactive elements (e.g: tooltip on hover, charts that re-render when you change the data criteria). You can see the github; and examples

datasette is another player in this space -- "data applications using sql" space, but evidence goes much further by supporting various data backends, and having first class support for thinking in terms of reports, and standard visualization components like charts. Support for markdown is naturally expected.

The combination of markdown + code does put it in a group with jupyter, but the availability of dynamic controls also make you think of streamlit, but does not go as far as observable because observable, given it's pedigree with d3 etc., is a lot more capable visulization tool (think maps, and advanced dataviz techniques).

Can we run AI and #ai tools offline? What would it take? These are interesting and important questions.

Some thoguhts on that -- I Want Everything Local — Building My Offline AI Workspace via instavm.io. In this article they talk about:

• #llm s run completely locally
• Code executes inside a lightweight #virtual-machine, not on the host machine
• headless browser for automation and internet access

-Fixie.ai

• turbopuffer

Pennsylvania is low key the king of American urbanism. It's like a time capsule, you've got hundreds of little towns and dozens of midsized cities where normal middle class people are still living in townhouses, shopping at corner shops, and walking for a daily errands. -- via

The title of this post is a good starting point to start thinking and capturing what would "AI Powered Data Engineering" look like? AI-Powered Data Engineering: My Stack for Faster, Smarter Analytics | Estuary. Tech mentioned: Roo as agentic framework, Claude Code (and CLAUDE.md), dbt for transformations, MCP servers, and duckdb. #phrases

GDPR-compliant European static site hosting - statichost.eu is a static hosting website that is an alternative to cloudflare and github pages.

It works by pulling your git repositories hosted elsewhere (github, gitlab, codeberg etc.,) and using a docker container to execute build scripts. it supports quite a few javascript based JAM stacks, and also hugo (golang), and zola (rust). It should be possible to create your own docker container and use it to build a site.

As an experiment, I put up AI Engineer website using hugo, with a private repo on github. In order to pull from a private repo, you can add an SSH public key provided by statichost as a Github Access Key. In order to trigger a build you can use the Webhook from github.

#static-sitegen

Litestar is a python framework that makes different set of choices compared to many popular frameworks likd fastapi, and flask, that help you in growing the application much better while not sacrificing speed. It also makes good use of #sqlalchemy while not being too tightly coupled with it. The routing system is also well thought out and helps in building a modular architecture.

See 2025-32.

#python #web-frameworks #fastapi