ʎɐʍollɐפ ʇʇoɔS

1.7K posts

ʎɐʍollɐפ ʇʇoɔS banner
ʎɐʍollɐפ ʇʇoɔS

ʎɐʍollɐפ ʇʇoɔS

@scottgal

MD: @[email protected] Email: [email protected] LI: https://t.co/R6V0bqDQEO I AM NOT PROF SCOTT GALLOWAY

It's in the trees! It's coming (also Scotland) Katılım Mart 2007
3.4K Takip Edilen3.1K Takipçiler
Sabitlenmiş Tweet
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
New website for the new venture stylobot stylobot.net A new type of bot detector. Mine is SUPER quick - sub millisecond Multi-Path - behavioural signature. Rolling window approach - decisons made during between ^ across requests safely, quickly. ZERO PII - NO PII at all (IP, user agent) saved to disk EVER Early days but the prototype github.com/scottgal/mostl… is functional. #botdetection #heuristic #business
ʎɐʍollɐפ ʇʇoɔS tweet media
English
2
0
2
195
ʎɐʍollɐפ ʇʇoɔS retweetledi
Dan Neidle
Dan Neidle@DanNeidle·
I see some weird things but this takes the biscuit. A vulnerability in the Companies House website, that let anyone view the private dashboard of any one of the five million registered companies, see directors' personal details. And modify them.
English
380
1.6K
7.4K
1.6M
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
Behavioural Inference Systems Most “AI coding” today means asking an code LLM to write your whole system for you. Behavioural inference systems flip that around: use code LLMs to build systems that infer *behaviour*. That 'behaviour' is really just metadata extracted from whatever entity it is you want to infer over. Be it http requests, documents, spreadsheets, data ,images etc..etc... Instead of one big model trying to understand everything, you generate specialised detectors, heuristics, and inference paths that observe behaviour and emit signals. The system learns how to combine them, refine them, and evolve them. This lineage runs from DiSE → Constrained Fuzziness → behavioural inference systems. Everythign from advanced structural Document Intelligence, Audience Segmentation, Bot detection, Hybrid RAG systems...all possible for this architecture. The goal isn’t smarter prompts. It’s building smarter systems with code-LLMs RAPIDLY which are (increasingly!) cheap to run and fault tolerant. mostlylucid.net/blog/behaviour…
English
0
0
0
94
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
I've low key invented a type of app, based on a "behavioural inference engine". ALL SORTS of possiblities for new systems, Spam, Segmentation, auto adjustng rate limits and accessible apis to clients etc... So now the 'poor cousin of OSS' .NET has a UNQUE application class 🤓 Think Audience Segmentation and analysis only...for web requests and in sub-millisecond timeframes for every request over MONTHS with hundreds of signals over 28 detectors, everything from UA to markov chain path analysis and behavioural drift detection. TINY (0.6b class...runs in CPU in 200ms!) LLM optional to describe stuff and handle edge cases. Free & Open Source Software stylobot.net #netcore #aspnetcore #aspnet #botdetection #behaviouralinference #foss
English
0
0
0
57
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
New blog article: StyloBot: How Bots Got Smarter - The New Frontier in Bot Detection (Part 2) mostlylucid.net/blog/botdetect… Covers in more detail how bot / scraper / fraud detection needs to change in the new post-LLM world of intelligent scrapers. And how my free OSS tool StyloBot moves the state of the art forward in respondign to these threats using a novel *no-LLM required* temporal behavioural signals based system. stylobot.net
English
0
0
0
39
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
v1 Of DoomSummarizer is out. It's a crazy deep research / auto knowledgebase system. Point it at a directory of word docs, pdf and markdown it'll index it all then answer questions about the contents. Point it at a url it'll parse the content, index it and tell you what it's about. Crawl your company's knowledgebase? It'll automatically become a support AI. Want to know what your biggest invoice was, when you sent that angry letter etc...all local, all private, all open source (unlicense) . Uses TINY 0.6b and 4b LLMs but can summarize / answer questions about whole books. #llm #ai #rag #search #localllm #ollama #onnx github.com/scottgal/lucid…
English
0
0
0
58
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
Deep Research without Deep Pockets. I pulled apart the premium “Deep Research” tools to see what they actually do, then built the same workflow locally without needing a huge model or huge spend. The trick: make the pipeline do the hard work (search + reduce + evidence), so the LLM mostly just writes. Part 5 of the DocSummarizer series: mostlylucid.net/blog/doomsumma… What’s your best technique for reducing “model made it up” without just throwing a bigger model at it? #rag #llm #deepresearch #ai #llm #lucene
English
0
0
1
31
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
🚧 DoomSummarizer (PREVIEW / ALPHA) I’ve been distilling the lucidRAG principles (hybrid search, entity extraction, knowledge graph + evidence-grounded synthesis) into a console-first, local-first research assistant + personal knowledge base. Think 'big boy's Deep Research running on a laptop with local llms'. It’s a CLI that can: Scroll: fetch + rank news/search into a digest / deep-dive / newsletter Ask: interactive Q&A over your stored evidence Crawl: index any site (incremental with ETags) Page: summarise a single URL Long-form: multi-section articles with grounding + validation MCP server: expose KB/search/entity graph to agents Runs fully offline after first model download (no API keys needed for default sources). Cloud LLM/search providers are optional + budget controlled. End of this post is DoomSummarizer… summarising its own README. 🙂 github.com/scottgal/lucid… doomsummarizer page "mostlylucid.net" --name doom doomsummarizer scroll "Tell me about DoomSummarizer" --name doom #ai #llm #csharp #rag #deepresearch #cli
ʎɐʍollɐפ ʇʇoɔS tweet media
English
0
0
0
27
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
Ok completing 4.1 and I'm DONE with ConsoleImage for a while. It can do live transcription of any video file to give synced subtitles. With color realtime rendering with a unique perceptual dithering system. NOW with scrubbing so you can use the cursor keys to navigate through movies! The slideshow mode lets you navigate through whole folders, viewing images and videos, animated gifs with live transcription of video *ALL IN THE TERMINAL*. Possibly the coolest thing I've ever built!. The image below is a screengrab of my Windows terminal playing realtime video. SHOULD be Linux & Mac compatible but LIGHTLY tested (appreciate bug reports!). Article: mostlylucid.net/blog/timeboxed… Tool: github.com/scottgal/mostl… #cli #video #terminal #tool #image #ai #subtitles #transcription
ʎɐʍollɐפ ʇʇoɔS tweet media
English
0
0
0
38
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
Time-boxed tool that escaped containment: ConsoleImage. A glyph-based terminal renderer using shape-matching, full-colour Braille (2×4 dots/cell), blocks, and temporal stabilisation so it’s actually watchable. Images, GIFs, videos, URL streams. Plus a custom document format and an MCP “visual probe” for AI workflows. mostlylucid.net/blog/timeboxed… (The image is rendered in terminal using an innovative braille based render). #dotnet #csharp #cli #terminal #ascii #braille #ffmpeg #opensource #devtools #ai #mcp
GIF
English
0
0
0
54
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
Problem: we keep using frontier LLMs as glue for jobs that are already solved. Solution: run OCR + NER locally in C# with ONNX Runtime. Deterministic extraction on ingest. Store the entities. Use an LLM later only if you actually need synthesis. OCR with Tesseract, then BERT NER via ONNX in .NET. No Python, no cloud, no tokens. This is my 'for beginners' article. I'm DEEP in OCR but realised I never explained the quickest way to do this *locally*. mostlylucid.net/blog/simple-oc… #CSharp #DotNet #ONNX #OnnxRuntime #OCR #NER #LocalAI #RAG #DocumentAI
English
0
0
0
41
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
LLMs are being used as sensors. That’s the mistake. In ReducedRAG, LLMs never see raw data. Deterministic pipelines extract facts first. LLMs only synthesize what’s already been reduced and verified. If your OCR, audio, or video pipeline starts with an LLM, you’ve already lost control. New article: Why LLMs Fail as Sensors (and What Brains Get Right) mostlylucid.net/blog/llms-fail… #ReducedRAG #AIArchitecture #LLMs #RAG #ComputerVision #MultimodalAI #SystemsThinking
English
0
0
0
43
ʎɐʍollɐפ ʇʇoɔS
ʎɐʍollɐפ ʇʇoɔS@scottgal·
Kinda sorta works. My little 1 day tool lucidVIEW. An Avalonia (so no Webview / bundled browser) Markdown viewer using Naiad (by @simoncropp , a super new library but for simple mermaid *it works*) to render svg (then to png because Skia.Svg SUCKS). Might go back to it but I just wanted something super lightweight (in runtime ;)) for stuff like this. Still ROUGH but 'tis done for the moment. github.com/scottgal/lucid…
ʎɐʍollɐפ ʇʇoɔS tweet media
English
0
0
0
64