DEV Community: gpt ai clips

OmniVoice: the free Apache-2.0 ElevenLabs alternative that runs on your own GPU

gpt ai clips — Fri, 05 Jun 2026 16:24:04 +0000

Two AI launches in one day

While most people were arguing about Claude Sonnet pricing, two genuinely useful AI projects shipped — and they both deserve a closer look.

1. OmniVoice — the open-source ElevenLabs alternative

OmniVoice is from k2-fsa, the team behind Next-gen Kaldi (Xiaomi's speech research lab). It's Apache-2.0 licensed, supports 600+ languages, and runs at RTF 0.025 — that's 40 seconds of audio generated per 1 second of compute.

Three modes:

Voice Clone — clone any voice with just 3-10 seconds of reference audio
Voice Design — generate a voice from controls (gender, age, pitch, accent, dialect)
Auto Voice — let the model pick

Install:

pip install omnivoice

Use:

from omnivoice import OmniVoice

model = OmniVoice.from_pretrained("k2-fsa/OmniVoice")
audio = model.generate(
    text="Hello from a free voice model",
    ref_audio="reference.wav"
)

CLI tools that ship in the box: omnivoice-demo, omnivoice-infer, omnivoice-infer-batch.

If you want to try it without installing anything, the HuggingFace Space lets you generate samples in your browser.

Why this matters: ElevenLabs is the de facto API for AI voice, but it costs ~$11/month minimum and 600+ language coverage is genuinely difficult to match. OmniVoice flips both: free + broader coverage + local-first.

2. Google DreamBeans — the anti-infinite-scroll AI

DreamBeans launched June 3, 2026 from Google Labs. It connects to your Gmail, Calendar, Photos, YouTube, and Search history, then uses Nano Banana 2 (Google's latest image gen) to brew a finite daily collection of illustrated stories about your life.

The interesting part isn't the tech — it's the product decision. Google explicitly chose NOT to make it infinite scroll. You get a fixed number of stories per day, and that's it.

For now it's US-only on iOS/Android for Google AI Ultra subscribers. Everyone else can join the waitlist.

60-second video walkthrough

Includes free PDF guide with all the install commands, prompts, and links.

More weekly AI deep-dives on @gptaiclips and the podcast.

Affiliate disclosure: my channel description contains a GoHighLevel referral link — I may earn a commission at no cost to you.

last30days-skill: the open-source AI agent that searches Reddit, X, HN, and Polymarket in parallel

gpt ai clips — Fri, 05 Jun 2026 09:07:50 +0000

27,000 GitHub stars in 30 days. What's the trick?

last30days-skill is an MIT-licensed AI agent skill that does one job extremely well: it queries six high-signal platforms at the same time and returns a single cited brief.

What it searches

Reddit
X (Twitter)
YouTube
Hacker News
TikTok
Polymarket (this is the unique angle — real money on the line)

Scoring

Results are weighted by upvotes, likes, views, and Polymarket odds. The agent synthesizes everything into one brief with citations you can verify.

Install in 10 seconds

# Claude Code
/plugin marketplace add mvanhorn/last30days-skill
/plugin install last30days

# Codex / Cursor / Gemini / 50+ others
npx skills add mvanhorn/last30days-skill -g

# OpenClaw
clawhub install last30days-official

Top prompts that just work

Use last30days to research [person] before my meeting
Use last30days: what just happened with [news topic]
Use last30days to compare [tool A] vs [tool B]
Use last30days: [geopolitical event] — include Polymarket odds
Use last30days to find everything on [trending model] this week

56-second video walkthrough

I'm covering AI agent skills weekly on @gptaiclips — podcast version on Spotify.

Affiliate disclosure: my channel description includes a GoHighLevel referral link — I may earn a commission at no cost to you.

Open Notebook Hit 25K Stars — The Self-Hosted, MIT-Licensed NotebookLM You Can Actually Own

gpt ai clips — Fri, 05 Jun 2026 06:10:29 +0000

The self-hosted NotebookLM alternative everyone is starring

Open Notebook just crossed 25,000 GitHub stars. It's MIT-licensed, 100% local, and supports 18+ AI providers including OpenAI, Anthropic, Ollama, and LM Studio. Drop in PDFs, YouTube links, audio, and web pages — get answers with citations. Generate multi-speaker podcasts with 1-4 voices. Your research data never leaves your machine.

Two-minute Docker install

# Grab the compose file
curl -o docker-compose.yml \
  https://raw.githubusercontent.com/lfnovo/open-notebook/main/docker-compose.yml

# Change OPEN_NOTEBOOK_ENCRYPTION_KEY to a long random string (32+ chars)

# Launch
docker compose up -d
docker compose ps
open http://localhost:8502

Ports: 8502 UI, 5055 REST API, 8000 SurrealDB (internal). Persistent data lives in ./surreal_data and ./notebook_data.

Three setups that actually work

Cheap & fast — Anthropic Claude Haiku. Best for citation-heavy long-context research.
Free + local — Ollama with llama3 or qwen2.5. Point Open Notebook at http://host.docker.internal:11434.
Best podcasts — OpenAI gpt-4o for the script, TTS-1-HD for multi-speaker audio. Configure 1-4 speaker profiles in Settings → Podcasts.

Real use-cases

Research → cited paper. 10-20 PDFs + a few web pages. Open-ended questions return cited answers. Beats ChatGPT for serious lit review.
Content production. Transcripts plus competitor blog posts in, outlines + multi-speaker podcast episodes out.
Legal review. Run it 100% on-prem with Ollama. NDAs, MSAs, policies — no client data ever crosses the network.

60-second walkthrough

I broke down the install + the dashboard in a short here: https://www.youtube.com/shorts/6WN52Mw3n28

The full audio breakdown lives on the GPTAIClips podcast on Spotify: https://open.spotify.com/show/033sVcOc9PNwH6eIX5SPSq

More short-form breakdowns of trending AI repos at https://gptaiclips.com.

Affiliate disclosure: I use GoHighLevel (Pro trial via that link) as the CRM I sell AI workflows out of — referral link, doesn't change your price.

Vibe-Trading Just Hit 10K Stars on GitHub — and It Actually Places Real Broker Orders

gpt ai clips — Thu, 04 Jun 2026 07:05:18 +0000

Vibe-Trading hit 10K stars overnight

A new open-source repo called Vibe-Trading just crossed 10,000 stars on GitHub, and the reason is unusual: it's an AI agent that actually places real orders across six brokers — Robinhood, Alpaca, Binance, OKX, Tiger, and one more.

Most "AI trading" projects stop at backtests. This one ships a real CLI, a real web dashboard, and connectors that authenticate against real brokerage APIs.

What you get

One-line install — pip install style setup, no Docker gymnastics
CLI loop — run the agent from your terminal, watch trades stream
Web dashboard — PnL, positions, model thoughts
6 broker connectors — across crypto and US equities

How to use it safely

The project ships with paper-trading mode on every broker that supports it. Turn that on first. Run the agent for a week with fake money. Read every trade in the dashboard. Only then consider letting it touch a small real-money account you can afford to lose.

This is open-source software placing real orders. Not financial advice. Read the code, understand what it does, and start with paper.

Watch the 60-second walkthrough

I broke down the install, the dashboard, and the broker list in a 60-second video here: https://www.youtube.com/watch?v=h97KknnguBo

More short-form breakdowns of open-source AI tools at https://gptaiclips.com — that's where I park the longer experiments.

GPT-5.6 Reportedly Spotted in OpenAI Codex Logs — Three Codenames, 1.5M Context

gpt ai clips — Thu, 04 Jun 2026 05:00:31 +0000

An identifier gpt-5.6 showed up in OpenAI Codex backend traces over the weekend, alongside three internal codenames — iris-alpha, ember-alpha, and beacon-alpha. Polymarket is currently sitting at 85%+ odds for a public release before June 30. Here is what was reportedly leaked, what to take with a grain of salt, and what it means if any of it ships.

The GPT-5.6 Leak

From Codex traces and a handful of corroborating Discord screenshots:

Field	Reported Value
Identifier	`gpt-5.6`
Internal codenames	`iris-alpha`, `ember-alpha`, `beacon-alpha`
Context window	1.5M tokens (+43% over GPT-5.5)
Tiers	Standard + GPT-5.6 Pro
Pricing rumor	2-3x cheaper than Anthropic Mythos at same tier
Focus	Agentic workflows + front-end generation

The three codenames suggest a routed-ensemble approach — different sub-models for different task classes — though that is interpretation, not leaked text.

The Claude Mythos Counter-Leak

Anthropic apparently ran Mythos against an 18-benchmark internal suite. Mythos won 17 of 18:

Benchmark	Mythos	Comparison
SWE-bench Verified	93.9%	Opus 4.6: 80.8%
SWE-bench Pro	77.8%	GPT-5.4: 57.7%
Terminal-Bench 2.0	82.0%	—
USAMO 2026	97.6%	—
GPQA Diamond	94.5%	—
BrowseComp	86.9%	—
Cybench	100%	saturated

Mozilla reportedly found 231 zero-day vulnerabilities testing Mythos against Firefox — 10x more than the previous Claude model could surface in the same harness.

The Microsoft Slip

A slide at Build 2026 briefly showed Mythos training compute at approximately 6.1 × 10²⁷ FLOPs — roughly 300x what GPT-4 was trained on. The slide was pulled from the recording shortly after, but several attendees screenshotted it.

How To Read All Of This

Codex backend identifiers are real signals — they have predicted releases before (GPT-4o, GPT-5-mini).
The Mythos scorecard is directional, not definitive — internal benchmarks are always optimized for internal context.
The Microsoft slide is the weakest signal — a single unverified screenshot.
Polymarket odds above 85% are meaningful — that market has been well-calibrated on OpenAI releases.

What To Watch

Any OpenAI API pricing page update showing a gpt-5.6 SKU
Anthropic's model card for Mythos (expected alongside release)
Codex UI changes in the next two weeks

This will resolve quickly. Neither company has a reason to delay past Q2 at this point.

All benchmark figures above are from leaked documents, not official model cards. Treat them as directional until vendor APIs or official releases confirm them.

Want more AI news like this in 60 seconds? Watch the 60-second walkthrough

If you want to stay ahead of every frontier-model leak, benchmark drop, and release signal, gptaiclips.com aggregates the signal from developer Discords, Polymarket, and API changelogs into one daily digest — free to follow.

I Pointed Claude Code at HackerOne and It Drafted a Valid Bug Report in 90 Minutes

gpt ai clips — Wed, 03 Jun 2026 07:04:50 +0000

Bug bounty hunting used to be a hand-cramping cycle of ffuf, nuclei, httpx, and a forty-tab browser. This weekend I wired all of it behind one Claude Code agent and walked away with a draft report queued for HackerOne triage. Here is the architecture, the install, and what I learned in the first 90 minutes.

What Claude Bug Bounty Actually Is

Claude Bug Bounty is an open-source agent framework that hands Claude Code five specialized sub-agents:

Sub-agent	Job
Recon	Subdomain enumeration, asset discovery, scope verification
Fuzzer	Endpoint discovery, parameter mining, auth-flow probing
Validator	PoC generation, severity scoring against CVSS 3.1
Reporter	Markdown draft tuned for HackerOne / Bugcrowd templates
Triage	De-duplicates findings against known CVEs and prior reports

The orchestrator decides which sub-agent to invoke based on the current state of the engagement. You drive it with plain English ("start with passive recon on this scope") instead of bash incantations.

60-Second Install

npm install -g @anthropic-ai/claude-code
git clone https://github.com/gptaiclips/claude-bug-bounty
cd claude-bug-bounty
cp .env.example .env  # add HACKERONE_API_TOKEN, BUGCROWD_API_TOKEN
claude

Claude Code picks up the .claude/agents/ directory automatically. Type /recon https://target.example.com and it begins.

The Important Ethics Block

This tool is only legal on programs that explicitly authorize testing. Use it on HackerOne / Bugcrowd / Intigriti / YesWeHack scopes you have been invited to, or against your own assets. Pointing it at a random site is unauthorized access — the framework intentionally refuses targets that are not in a recognized program list.

What 90 Minutes Got Me

47 subdomains across the in-scope program
3 candidate IDORs (2 confirmed via the Validator sub-agent)
1 draft report at Medium severity (CVSS 6.5), submitted for triage
Payouts in this program: $250 (Low) → $15,000 (Critical)

The Reporter sub-agent's draft was good enough that I made two wording edits and submitted. That is the part that surprised me — the report quality is what usually delays a payout, not the finding itself.

Full Walk-Through

The long-form video covers recon, fuzzing, validation, and reporting end-to-end, plus the payout math for HackerOne vs Bugcrowd vs Intigriti:

Free install guide (.env template, scope policy, agent prompts): gptaiclips.com/go/ep59

What is your first target program going to be? Drop it in the comments — I am tracking which scopes are most agent-friendly.

I Stacked 3 GitHub Repos Into a Weekend AI Services Business

gpt ai clips — Sun, 17 May 2026 09:48:09 +0000

Three GitHub repos crossed 2,500 stars this month and — by accident — they slot together into a deployable AI services stack. Here is how I am running them on one terminal.

The Three-Layer Stack

Layer	Repo	Stars	What It Does
Visual	neilsonnn/image-blaster	2,944	Photo → explorable 3D world in ~5 minutes (Claude Code × World Labs marble-1.1 × FAL hunyuan-3d × ElevenLabs SFX)
Front-end	nexu-io/html-anything	2,599	Agentic HTML editor — auto-detects 8 coding CLIs, 75 skills × 9 surfaces, zero API keys required
Memory	Tencent/TencentDB-Agent-Memory	2,530	4-tier local memory (L0–L3), v0.3.4 released May 13 2026

Why the Memory Tier Matters

Tencent's published benchmarks for the L0–L3 hierarchy versus flat RAG:

Metric	Flat RAG → 4-Tier
WideSearch	33% → 50%
Tokens used	−61.38%
PersonaMem	48% → 76%

That is roughly half the tokens for double the persona recall — which is the difference between a demo and a paid product.

Install in One Terminal Session

git clone https://github.com/neilsonnn/image-blaster && cd image-blaster && bun install
git clone https://github.com/nexu-io/html-anything && cd html-anything && pnpm install
git clone https://github.com/Tencent/TencentDB-Agent-Memory && cd TencentDB-Agent-Memory && uv sync

No glue code required — html-anything auto-detects whichever CLI is on your PATH, and TencentDB-Agent-Memory exposes an HTTP shim that image-blaster can call as a memory backend.

Monetization Angles

Visual-first agency — sell branded 3D walk-throughs (real estate, product launches) using image-blaster + html-anything landing pages.
Long-running personal agents — license a private assistant with the 4-tier memory that actually remembers clients between sessions.
White-label AI dashboards — html-anything's 75 skills × 9 surfaces map onto every common SaaS dashboard pattern.

Walk-through (≈60 seconds):

More stack breakdowns: gptaiclips.com · prev episode: gptaiclips.com/v/ep57

Give Your AI Agent a Real Memory in One Command (Hermes v0.14 + Obsidian)

gpt ai clips — Sun, 17 May 2026 00:34:46 +0000

Most AI agents have amnesia. You explain the same context every session. Here is a five-minute fix that turns any LLM into something with actual continuity.

The trick: use a markdown vault as memory

A vector database is the obvious choice and usually the wrong one. Embeddings give fuzzy recall and lock you into a specific provider. A plain Obsidian vault gives the agent exact, named, editable memory that the human can also read.

Hermes Agent v0.14 shipped May 16, 2026 with a first-class Obsidian provider. One command wires it up.

Install Hermes

Three paths, pick one:

pip install hermes-agent && hermes postinstall

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

irm https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.ps1 | iex

Wire in Obsidian

hermes memory setup --provider obsidian --path ~/vaults/work
hermes memory status

That is the whole memory setup. The agent now reads and writes markdown files in that vault every session.

Optional power-ups

Obsidian Local REST API plugin on localhost:27123 gives programmatic R/W during agent execution, not just at boot.
Obsidian MCP server (cyanheads/obsidian-mcp-server) exposes the vault as Model Context Protocol tools — any MCP-aware client can use the same memory.
Omi captures passive screen and mic context into the vault, which turns the workflow into a true second brain.

What else landed in v0.14

22 messaging platforms (Telegram, Discord, Slack, WhatsApp, Signal, Teams, LINE, more)
400+ supported models (xAI Grok SuperGrok OAuth, Nous Portal, OpenRouter, Ollama)
68 built-in agent tools
hermes proxy — OpenAI-compatible local endpoint for OAuth-only providers
/handoff — live session transfer between devices
19s faster cold start, 180x faster browser CDP
MIT license, 100K+ stars (fastest project to 100K)

60-second walkthrough

Why this matters for builders

If your agent stack is a glue script that pastes the same context into every prompt, you are paying tokens for amnesia. Move the context out of the prompt and into a vault, and prompts shrink while quality goes up. Long breakdowns and reference vaults live at cptdigital.com.

From abandoned repos to a $87K Obsidian vault: a three-pass extraction pattern

gpt ai clips — Sat, 16 May 2026 11:14:40 +0000

Most of us have a folder full of repos that haven't been opened in months. Here is a pattern I have been using to turn those repos into a packaged developer product instead of guilt.

The three-pass pipeline

The core idea is to extract decisions, not descriptions. A file-level summary is useless to a future reader — what they want is the implicit reasoning the original author was preserving.

Pass 1 — file-level extraction

For every file, ask the model for four things: purpose, public surface, hidden invariants, and a risk score from 1 to 5. The risk score is the secret ingredient — it forces the model to find load-bearing logic.

Pass 2 — module-level clustering

Feed all of pass 1 back in and ask for clusters of files that share invariants. Each cluster becomes an Architecture Decision Record (ADR) with status, context, decision, and consequences.

Pass 3 — architecture-level graph

Apply Leiden clustering across the ADRs to surface the cross-cutting concepts. Each graph node carries a one-paragraph 'why this matters to a maintainer' note. Leiden gives stabler cluster boundaries than vanilla modularity on small graphs.

Why Sonnet 4.6 matters here

With a 1M-token context window I can run the whole-repo pass without first summarizing per file. Per-file summarization is where cross-references die — once you compress, you lose the links the graph step depends on.

Packaging into Obsidian

Graphify (the Claude Code skill, ~37K stars) has an --obsidian flag that writes the graph as a markdown vault with backlinks already wired up. Add ADR templates and you have a product, not a dump.

Quick demo

Three-tier pricing

$0 — sampler vault with two ADRs and the graph view
$49 — full vault with all ADRs and concept notes
$149 — full vault plus raw prompts and ADR templates so the buyer can run the pipeline on their own code

Top developer vaults on Gumroad clear 3K+ copies a year, so the ceiling is real.

Try it on your own repo

Checkout the Graphify project and the longer walkthroughs over at cptdigital.com. The bottleneck is almost always the prompt for pass 1, not the model.