sean-hidock

Everyone argues over which AI is smartest. The scarier question: where does your data go the second you use one? The tools that win the next decade won't be the smartest — they'll be the ones where your private stuff never leaves your machine. Own your inputs; the rest is rented.

@xenovacom The part I find wild: for years "runs locally" meant "runs a toy." 1-bit at 27B-class flips that — local is no longer the compromise tier. The interesting fallout isn't just browser demos; it's which apps get rebuilt now that private-by-default stops costing you capability.

Bonsai 27B just changed the local LLM game forever. 1-bit quantization shrinks it from 54GB to just 3.8GB (-93%), while retaining 90% of its intelligence. That's insane. With custom WebGPU kernels written by Fable 5 and GPT 5.6 Sol, the model now runs locally in your browser!

@PrismML This is the unlock the privacy conversation was waiting for. "Keep your data local" always carried a hidden tax — local meant weaker. A 27B model on the phone collapses that. Once on-device matches the cloud, there's no reason left to hand your private stuff to a server.

Today, we’re announcing Bonsai 27B: the first 27B-class model to run on a phone. Bonsai 27B is the new multimodal flagship of the Bonsai family. Based on Qwen3.6 27B, it brings a new capability tier to local AI: multi-step reasoning, structured tool use, long-context workflows, and coherent agentic loops. Until now, models in this class have been impractical to deploy locally. A 27B model occupies roughly 54 GB in 16-bit precision, and even a strong 4-bit build is around 18GB - too large for a phone and for most laptops. Bonsai 27B changes that. It comes in two variants: • Ternary Bonsai 27B: 5.9 GB, 1.71 effective bits per weight, optimized for laptop-class quality. • 1-bit Bonsai 27B: 3.9 GB, 1.125 effective bits per weight, optimized for phone-class footprint. Everything is open-sourced today under the Apache 2.0 license.

@biggerztrends The backlash worked this time — but look at the default it exposed: your public photos were fair game until enough people objected. "Pull it after outcry" isn't consent, it's damage control. The real fix is tools built so your stuff was never theirs to train on.

JUST IN: Meta has pulled its new "Muse Image" AI tool that turned public Instagram photos into AI-generated images just days after launch, following backlash over consent and privacy.

@nickgomez @openknowledge "Local and private" as the default for a 2nd brain is exactly right — one you can't inspect or export was never really yours. The half still under-built everywhere: making retrieval legible, so you see WHY something surfaced, not just trust it did. Rooting for this.

Introducing @OpenKnowledge, the best markdown IDE for humans and agents. Open source. Local and private. LLM-wiki ready. Use with Claude, Codex, and your favorite agent today.

@AlexFinn The capability's the headline. The quieter part matters more: when the model runs on your desk, your private stuff — meetings, notes, calls — never leaves it. Local isn't just faster. It's the first time "private by default" is physics, not a policy you have to trust.

You are going to be able to run Fable 5 locally on your desk In 2 years Apple will be coming out with Mac Studios with 1.5TB of memory With just 300gb of memory you can run Opus 4.8 level intelligence Think of what you can do with 5x that You need to be preparing for this now Start getting familiar with local AI technology Go to your Hermes/OpenClaw and use this prompt: “I am brand new to local AI and want to get familiar. Look at the computer you are currently on. Understand the specs. Then go on Huggingface and find the best models I can run on it. Then, walk me through how these models work, how they will run locally, and use cases I can do with them. After walking me through all of that so I’m educated, you can then load it onto this computer and build an interface so I can use them” In 2 years EVERYONE on Earth will have a local model running on their desk The people who start preparing now will be WAY ahead of everyone else

The trap isn't that AI is evil — it's that "free" and "convenient" quietly cost you your data the moment you type. The fix isn't a better privacy toggle. It's tools where the sensitive stuff — meetings, notes, calls — is handled where it's made and never leaves.

Everyone argues over which AI is smartest. The scarier question: where does your data go the second you use one? The tools that win the next decade won't be the smartest — they'll be the ones where your private stuff never leaves your machine. Own your inputs; the rest is rented.

People ask "which AI should I use?" The better question: "which one can I audit?" If you can't see why it answered — or where your data went — you're not the user. You're the training set.

@ayakacoop HiDockを作っている者です😊 「小さすぎ」まさにそこを目指しました。スト6のボイチャ練習に使うの、めっちゃ良い発想ですね笑 会議もゲームも、録っておくとあとから文字起こしで振り返れるので、使ってみて気になった点があればぜひ教えてください！

会議の声録音できるの小さすぎ…仕事後にスト6のボイチャ練習にもいいかも HiDock P1 mini 本体 ボイスレコーダー jp.mercari.com/item/m39297228…

@XFreeze "Opt-out" is the trap — it makes escaping a default you never chose your job, and you still can't verify it worked. The real fix isn't a better opt-out; it's tools where there's nothing to opt out of, because your data was never collected. Architecture beats fine print.

A reminder: Across much of the AI industry, user data is retained by default unless you manually change the privacy settings and even then, opt-outs often apply only to future data, may be limited by exceptions buried in the legal fine print, and can be difficult to verify in practice In other words, “opting out” does not always mean your data is fully deleted or excluded...the company’s own way of saying opt-out may not actually match what users reasonably expect in the end However SpaceXAI just showed what going beyond the industry standard actually looks like Running "/privacy" in Grok Build to disable retention also deletes all previously synced data...not just data created after changing the setting and prevents future code and session data from being retained Elon also confirmed that, as a precautionary measure, all user data uploaded to SpaceXAI before now will be completely deleted, with nothing remaining Yet some people are having a selective outrage meltdown, acting as though AI companies retaining uploaded data is some completely new revelation Data retention has long been common across major AI products, while manual privacy opt-outs are often buried, limited by exceptions, or focused primarily on future activity Meanwhile, SpaceXAI is setting a new privacy standard that goes far beyond what the rest of the AI industry currently provides by default

@Hesamation The "how big can I run locally" race sort of misses the point. For a personal assistant, a 1T model that doesn't know your week loses to a modest one that does. The unlock was never model size — it's getting your private context in without shipping it to a cloud first.

Apple’s M7 Ultra chip rumored to be a local AI workhorse as it supports 1.5TB of RAM. for reference, that’s enough to run Kimi or GLM models locally.

@kaz_fukumaru 取り上げていただきありがとうございます😊 HiDockを作っている者です。イヤホンをつけたまま会議もオンライン通話も録れて、あとから文字起こしで振り返れるところが、実際に使うといちばん効いてきます。使ってみての感想もぜひ聞かせてください！

Mac miniと相性抜群のAIオーディオコンパニオン「HiDock H1 Lite」もAmazon Prime Dayセールでお得に購入可能です！ 高品質なマイク・スピーカー機能を拡張でき、会議音声の録音と文字起こし・要約の生成もできます😊 お気に入りのワイヤレスイヤホンを使用した状態での録音にも対応しています！

@interesting_aIl This is the tell: "consent" that's really coercion — train on your data or lose it. It doesn't have to be this way. Software can process what you give it and keep it yours; harvesting is a business choice, not a technical need. Pick tools that don't make you the training set.

Samsung Health is warning users that their data could be deleted if they refuse to let Samsung use it to train AI

@RoundtableSpace The underrated line is "citing your own pages instead of guessing." That's the whole game — an answer you can trace to the source beats a confident one you can't. Plain markdown wins because the trust is inspectable: you see WHY it surfaced something, not just that it did.

Andrej Karpathy, the CEO of Obsidian and Claude Code just built the smartest second brain on earth. The frame is simple. Obsidian is the IDE, Claude Code is the programmer, your notes are the codebase. Three commands run the whole system. → Ingest splits any article, podcast or PDF into atomic pages linked to everything you already know → Query answers from your own notes in your own words, citing your own pages instead of guessing from training data → Lint walks the entire vault weekly, kills stale claims and rewires orphan notes back in Then the Obsidian CEO shipped 5 skill files that teach Claude to write Obsidian's native language natively. The repo hit 41,000 stars. No vector database, no embeddings, no memory app. Just a folder of plain markdown and an agent that never gets tired of the linking, filing and upkeep that killed every note system before it.

@AskDeepshi "Before it's too late" is the whole thing. Privacy with AI isn't a setting you flip later — it's decided at input time. Once a private conversation reaches someone's model, it's in the training set for good. The good version keeps sensitive inputs from ever leaving.

Hard not to be pessimistic about most AI, given how they use your data. But private, unfiltered AI can still do a lot of good … if we don’t wait until it’s too late. x.com/effthealgorith…

@ElianaBlythe_Z Right — and "in the rush to adopt" is the trap. The exciting features get all the attention; the boring data-ownership clause is the one that quietly decides what happens to your conversations later.

@seansong Such a critical point about data sovereignty that often gets overlooked in the rush to adopt new tools.

When an AI note-taker gets acquired, your old recordings go with it — new owner, new privacy policy, same conversations, and you never get a vote. "Where does my data go if you're bought?" is the question worth asking before you hit record, not after.

@DamiDefi "A retrieval system, not a database" — that's the whole game. One axis it skips: what you'll feed it. The highest-value inputs (meetings, private docs) are exactly what people won't send to the cloud. A second brain's only as good as what you trust it with.

EVERYONE HAS BEEN TALKING ABOUT AI SECOND BRAINS. BUT ALMOST EVERYONE IS BUILDING THEM THE WRONG WAY. Most people think an AI second brain is just dumping notes into a folder. It's not. The real advantage isn't storing more data. It's designing a system your AI actually knows how to navigate. Here's the 5-level framework that changes everything: Level 1: Build routing before intelligence. Give your AI clear instructions on where every type of information lives, so it stops asking you the same questions repeatedly. Level 2: Turn isolated files into connected knowledge. Organize notes into topic-based wikis that let AI traverse concepts instead of opening random documents. Level 3: Add semantic search only where it matters. Keyword search finds words. Semantic search finds meaning. But don't vectorize everything because some data still needs full-document context. Level 4: Map relationships, not just documents. Knowledge graphs let AI understand how people, projects, companies, and decisions connect instead of treating them as separate files. Level 5: Automate the brain, not the chaos. Continuously syncing everything sounds powerful until irrelevant data becomes permanent noise. Curate what enters your second brain. The biggest insight? Your second brain isn't a database. It's a retrieval system. The winners won't have the most information. They'll have AI that knows exactly where to look, in what order, and when to stop searching. That's the difference between an AI assistant and an AI operating system. Video cred: Nate Herk | AI Automation (Youtube)

@rileywestreel Fair — "inherently" was too strong. Narrower point: with flat files the failure mode is legible by default, no tooling required. Vectors get there too, but you're buying the observability layer you described. For small KBs that maturity tax isn't worth it yet.

@seansong True, but vector stacks aren't inherently opaque. You can log retrieval scores and inspect embeddings too. The real gap is tooling maturity, not the approach itself.

@bouncy_news 取り上げていただきありがとうございます😊 HiDockを作っている者です。イヤホンをつけたまま会議もオンライン通話も録れる手軽さは、実際に使うと効いてきます。あとから文字起こしで振り返りもぐっと楽になるので、気になった方はぜひ。

イヤホン越しの会話もラクラク録音！ AI文字起こし対応のボイスレコーダー「HiDock P1」 by HiDock