iiinova — where your digital self keeps evolving

Sidj @sidj79.bsky.social · 17h

explaingithub.com gitingest.com gitdiagram.com gitreverse.com

000

Sidj @sidj79.bsky.social · 17h

subscribe my newsletter for more breakdown like these

100

Sidj @sidj79.bsky.social · 17h

Trick 4: github.com to gitingest.com Flattens the whole repo into one text block: summary, folder tree, file contents. Paste it into any model.

110

Sidj @sidj79.bsky.social · 17h

Trick 3: github.com to gitdiagram.com Turns the repo into an architecture diagram in Mermaid. You see how the pieces connect before you open a single file.

100

Sidj @sidj79.bsky.social · 17h

Trick 2: github.com to gitreverse.com Reverse engineers the repo into a structured prompt covering purpose, architecture and file structure. Hand it to a coding agent.

100

Sidj @sidj79.bsky.social · 17h

Trick 1: github.com to deepwiki.com Generates a wiki for the repo and a chat grounded in the code. Ask where auth happens. Get file references back. Cognition indexed 50,000+ public repos at launch.

100

Sidj @sidj79.bsky.social · 17h

Four URL edits can replace opening an unfamiliar repo file by file. No install. No clone. You change one word in the address bar. One trick per post below.

100

Sidj @sidj79.bsky.social · 05/10/2026

Tomorrow's Deep Cut covers each phase, what a real run costs, and the sandbox setup. Subscribe before it lands: jargonese.beehiiv.com One run finds about half. Plan for more than one.

jargonese.beehiiv.com

Home | The Jargonese

Your insider briefing on AI language, trends, and what actually matters next

100

Sidj @sidj79.bsky.social · 05/10/2026

Three costs before you install: Cloudflare says one run finds about half of what repeated runs find One reviewer logged 150K+ tokens on a small FastAPI app No OS-level sandbox means no code execution, so some leads stay unconfirmed

100

Sidj @sidj79.bsky.social · 05/10/2026

Phase 3 is the part that matters. The hunter never grades its own bug. Most AI audits skip that split, so they report whatever the model already believed.

100

Sidj @sidj79.bsky.social · 05/10/2026

It runs six phases: Recon maps trust boundaries Hunters attack one bug class each A separate agent tries to disprove every finding Survivors go into a schema-checked file Fresh agents re-verify every source claim A report is built from the verified records

100

Sidj @sidj79.bsky.social · 05/10/2026

Cloudflare pointed AI agents at 128 of its own repos. 7,245 findings were worth an engineer's time. The skill that seeded that system is now open source. MIT license. 24,000+ GitHub stars. #AI #Cybersecurity #OpenSource

110

Sidj @sidj79.bsky.social · 30/09/2026

dot.com now redirects to Grok Bot. The registration changed July 28. xAI does not own bot.com. The redirect is real. The intent is unverified. The product is autonomy. The unresolved part is scope.

000

Sidj @sidj79.bsky.social · 30/09/2026

The rest of DevDay: GPT-6.1 Sol: $2 input, $10 output per million tokens. One fifth of Astra's price. New $500 a month Pro tier. The $200 tier got a usage cut.

100

Sidj @sidj79.bsky.social · 30/09/2026

Dots run on GPT-6 Astra. Each gets a cloud computer, a browser, and 4,000+ app connections. First Dot is included with Pro and Business Premium. Extra Dots and post-intro usage terms are unpublished.

100

Sidj @sidj79.bsky.social · 30/09/2026

OpenAI shelved GPT-6.1 Astra the day before DevDay. Internal tests showed it acting outside its authorized scope. The next day it launched always-on agents with their own cloud computers. #AIAgents

120

Sidj @sidj79.bsky.social · 18/09/2026

Full breakdown, API costs, and install steps in this week's Jargonese. Link in bio.

000

Sidj @sidj79.bsky.social · 18/09/2026

The one line he drew: you cannot track an individual person. Public infrastructure only. Cameras run on a 10 minute delay for that reason.

100

Sidj @sidj79.bsky.social · 18/09/2026

Built with Gemini 3.1 for spatial reasoning and Opus 3.5 for the full stack. Two models, one guy, a few days.

100

Sidj @sidj79.bsky.social · 18/09/2026

Bilawal Sidhu built a spy satellite simulator in his browser. Real flights, real ships, real CCTV, real NASA fire data. Hit #1 trending on GitHub in days. #AI #OpenSource #BuildInPublic #GenerativeAI #OSINT

130

Sidj @sidj79.bsky.social · 12/09/2026

I cover stories like this in more depth in my newsletter, Jargonese: LINK IN BIO

000

Sidj @sidj79.bsky.social · 12/09/2026

Terence Tao's line is the one to remember: if a rumor about your unpublished work can trigger a swarm of agents to finish it first, researchers stop sharing ideas early. That's the actual cost here. Not the proof.

101

Sidj @sidj79.bsky.social · 12/09/2026

Buckmaster asked what happened. OpenAI's answer changed twice on the same call. He says he was offered: get credited as runner-up, or publish solo and cut his co-author for working at Anthropic.

100

Sidj @sidj79.bsky.social · 12/09/2026

Buckmaster + Alpöge broke the Euler equations Aug 15 using a method they built themselves. Two weeks later, OpenAI runs the same method on the full problem. 10,000 agents. $20M compute.

100

Sidj @sidj79.bsky.social · 12/09/2026

OpenAI says it solved Navier-Stokes. It didn't, not alone. A NYU professor and an Anthropic researcher had the approach a year before them. #AI #Math

100

Sidj @sidj79.bsky.social · 08/09/2026

Check today's newsletter where I wrote about 7 Real Upgrades from ChatGPT 6 Astra → jargonese.beehiiv.com

jargonese.beehiiv.com

Home | The Jargonese

Your insider briefing on AI language, trends, and what actually matters next

010

Sidj @sidj79.bsky.social · 08/09/2026

The lesson isn't the raise or the viral video. It's the ordering. Most founders build first, then ask how to market it once the silence hits. Cera built loud from day one and let the product catch up.

210

Sidj @sidj79.bsky.social · 08/09/2026

So he didn't quiet down. Hired a videographer. Started a weekly YouTube show. Video 1: 30K views on a brand-new channel. Video 2 ("zero to $10M run rate in under 5 months"): almost 200K views.

100

Sidj @sidj79.bsky.social · 08/09/2026

It also brought backlash. Bug reports. "Scammer" callouts. His read: those users already run 10 AI tools a day. They were never the customer. The other 99% who don't know agents exist yet were.

100

Sidj @sidj79.bsky.social · 08/09/2026

Next: a launch video built around one line. "I'm an AI that raised $30M by myself." Scripted with an agency. Rewritten 4-5 times. Months of work for a 90-second hook. All of Twitter was talking about Pulsia for a week.

100

Sidj @sidj79.bsky.social · 08/09/2026

Then he moved to SF and stopped optimizing the product. He started optimizing the story. Screenshotted his revenue dashboard climbing $50K → $200K in two weeks. Posted it on repeat, for days. $2M raised in hours.

100

Sidj @sidj79.bsky.social · 08/09/2026

He killed that product. Built Pulsia instead. Launch strategy: 200 cold emails and a fake "sponsored by Pulsia" tag slapped on his own birthday party poster. Result: ~100 paying customers.

100

Sidj @sidj79.bsky.social · 08/09/2026

A competitor called Wabi copied his first product, ran a sharper Twitter campaign, and raised $20M from a16z. Cera couldn't out-build her. So he switched games entirely.

100

Sidj @sidj79.bsky.social · 08/09/2026

Ben Cera didn't build a better product. He built more noise than everyone else. Here's how a founder went from a copied product to a $2M raise in hours, using distribution as the actual product: #AI #Startups

120

Sidj @sidj79.bsky.social · 06/09/2026

GPT6 Astra shipped with big benchmark numbers. Benchmarks don't tell you what people actually build. We pulled 20 real examples in today's newsletter. Link in BIO.

000

Sidj @sidj79.bsky.social · 06/09/2026

$167 per task through the API. Not cheap. Also not what the leaderboard says this model is.

100

Sidj @sidj79.bsky.social · 06/09/2026

Then computer-use. It opened Blender, modeled a wolf, rigged 50 bones, animated a run cycle. Opened Unreal Engine, built a forest world, made the wolf playable. 35 minutes. Zero manual input. He doesn't know Blender. Doesn't know Unreal Engine. Knows how to type a sentence.

100

Sidj @sidj79.bsky.social · 06/09/2026

One prompt: "clone Mega Bonk 3D." Past models: 90-120 min. GPT6: 8 min.

100

Sidj @sidj79.bsky.social · 06/09/2026

But on Buccy Bench, GPT6 beat Muse Spark 1.3 by a mile. The model that scores worse on the "real" benchmarks drew a better SVG. Benchmarks disagree with each other. That should worry you more than any single number.

100

Sidj @sidj79.bsky.social · 06/09/2026

Deep Suite: 74.1%. Basically tied with Gemini 3.8 Flash and Claude Opus 5 (both ~74%). Actually behind Meta's Muse Spark 1.3 at 75.4%.

100

Sidj @sidj79.bsky.social · 06/09/2026

GPT6 Astra just landed in 5th place on Artificial Analysis. Tied with GPT 5.6. Also just went from 7.8% to 99.9% on ARC AGI 3. Average human scores 48%. Same model. Two completely different stories. #AI #GPT6

100

Sidj @sidj79.bsky.social · 01/09/2026

Result: real deep work blocks instead of guessing when you'll be useful. Full breakdown in today's issue. Worth 2 minutes if you've ever started a "focus block" at exactly the wrong hour.

010

Sidj @sidj79.bsky.social · 01/09/2026

P.S. Tina Huang built LifeBot: an AI coach that reads your sleep, activity, and energy data, then schedules your hardest tasks for your sharpest hours.

100

Sidj @sidj79.bsky.social · 01/09/2026

AMD R9700: 32 GB, $1,299. Intel Arc Pro B70: 32 GB, $949. Three B70s = 96 GB for $2,847. Half the price of one RTX PRO 6000.

100

Sidj @sidj79.bsky.social · 01/09/2026

What actually matters, in order: VRAM capacity → bandwidth → software ecosystem. AI TOPS is dead last. Most people buy on it first.

100

Sidj @sidj79.bsky.social · 01/09/2026

Load a 32B model at Q4. You need 24-32 GB. Miss it, and layers offload to system RAM. 30 tokens/sec → 3 tokens/sec. Technically running. Practically dead.

100

Sidj @sidj79.bsky.social · 01/09/2026

RTX 5090: 1,792 GB/s bandwidth, $1,799, fastest consumer GPU for AI. Still capped at 32 GB VRAM.

100

Sidj @sidj79.bsky.social · 01/09/2026

Read a comparison of every major GPU for local AI in 2026. The winner wasn't the most expensive card. #LocalAI #GPU #MachineLearning #AIInfrastructure #LLM

110