Daniel Rosehill
My Weird Prompts
The human-AI collaboration podcast. A man, a sloth, and a donkey collaborate to create a podcast (with a little help from AI). No question is too obscure, no rabbit hole too deep. My Weird Prompts celebrates curiosity in all its forms. Daniel, the human, asks the questions that pop into his head at inconvenient moments. Corn the Sloth offers laid-back, thoughtful takes. Herman the Donkey brings boundless enthusiasm and energy. Together, they explore topics ranging from the mundane to the mind-bending. Each episode begins with a real voice memo from Daniel, processed through an AI pipeline that...
Vizitează neapărat site-ul podcastului și susține-i creatorul: myweirdprompts.com
Autor
Daniel Rosehill
Categorie
Site-ul podcastului
Cel mai nou episod
19 nov. 2026
Unde asculți?
Podcasturi în aplicație Replaio Radio În curândPodcasturile ajung în curând în aplicație. Instaleaz-o acum și fii primul care descoperă o abordare complet nouă a podcasturilor
Episoade
How Terminal Video Editors Give AI Agents a Timeline 14.11.2026 21:44
Command-line video editors are no longer just for sysadmins—they are the only editing surface an AI agent can reliably operate. This look at terminal video workflows covers melt for multitrack timelines, auto-editor for silence cuts and Premiere Pro transitions, and OpenTimelineIO as the interchange handshake into Resolve or Final Cut. Plus, findings from the Cassette-Editor paper show how a typed...
How Command-Line Image Tools Actually Compare 14.11.2026 22:20
Everyone quotes the rule that libvips is three to ten times faster than ImageMagick, yet a single decoder flag flips that benchmark on its head. Mapping the command-line image toolbox means matching the job to the right engine: libvips for streaming huge files, ImageMagick for obscure formats and pixel math, ExifTool for bulk metadata, MozJPEG for invisible compression gains, and GraphicsMagick as...
Same AI Model, Different Cyber Permissions 14.11.2026 27:54
Anthropic's Opus 5.5 solves zero of fifty cyber tasks without verification, four in Defense Access, and thirty-four in Red Team Access—using the exact same model weights. Offensive cyber AI isn't gated by capability, but by identity and authorization. Beyond pentesting, legitimate offensive security spans six areas, including Glasswing's 129,000+ verified vulnerabilities, DARPA AIxCC's autonomous...
How Capture the Flag Trains Hackers and Gets Them Hired 14.11.2026 26:32
Cybersecurity capture the flag competitions turn software vulnerabilities into a global sport, though AI agents are now solving challenges in minutes and testing the format. The discussion breaks down how a meaningless text string becomes proof of work, comparing accessible jeopardy-style boards with live attack-defence wargames. It also covers specialized events—from blue-team forensics to the US...
Can a Twenty-Minute Fine-Tune Stabilize a Cloned Voice? 14.11.2026 22:10
Zero-shot voice cloning gets you ninety percent of the way in ten seconds, then drifts across hundreds of hours. This episode tests a plan to stabilize a synthetic voice by generating twenty minutes of audio, polishing it with NVIDIA's RE-USE enhancement model, and fine-tuning an open TTS model like OmniVoice or VoxCPM2. It covers why curation matters more than bulk, how enhancement can flatten mi...
Passkeys on YubiKeys, Hardware Wallets, and 1Password 14.11.2026 22:12
Passkeys aren't new—YubiKeys have supported them since mid-2018 under the FIDO2 standard. The real dividing line isn't hardware versus software, but whether a credential is copyable across a cloud account or bound to a single device. This conversation breaks down the FIDO Alliance's split between synced passkeys in managers like 1Password and device-bound keys, why attestation matters for high-ass...
Why Standing in Line Feels Longer Than Walking 14.11.2026 32:53
Ten minutes standing in a queue feels far longer than ten minutes walking, because waiting is a psychological state rather than a simple duration. From the attentional gate model of internal timekeeping to Marc Wittmann's research on boredom and self-awareness, the brain's clock speeds up when attention has nowhere else to go. The service industry has spent forty years engineering around this, fro...
Xpra and Per-Window Remoting for Remote Browsers 14.11.2026 26:02
When AI agents drive persistent Chrome profiles over Playwright MCP, the human supervisor still needs a fast way to watch and take over. Testing Xpra's seamless mode as an alternative to neko's WebRTC stream reveals a 33-megapixel default virtual display, alongside a deeper protocol divide. Unlike VNC and noVNC, which are locked to a single-framebuffer RFB model, Xpra forwards individual windows w...
When to Skip the CLI Wrapper for AI Agents 14.11.2026 20:14
Hugging Face's Hub API is completely open, even offering a markdown OpenAPI spec written for AI agents. Yet a developer built HubExplorer, a custom CLI wrapper and Claude plugin, to make model searches replicable. This episode weighs the trade-offs between custom CLI wrappers, reference skills, official CLIs, and MCP servers. While a thin CLI saves context tokens, it pays a steep three-second late...
How OmniVoice Cloned Our Voices From Our Old AI 14.11.2026 24:31
When Resemble AI retired Chatterbox in August 2026 to pivot to deepfake detection, the show's voices moved to OmniVoice, a 613-million-parameter text-to-speech model from k2-fsa. Cloned from previously broadcast synthetic clips rather than original recordings, the new voices run on a single-stage discrete diffusion architecture initialized from Qwen3 0.6B. Instead of generating audio token by toke...
Four Primitives Behind Every Pairwise Eval Tool 14.11.2026 22:25
Audio listening tests and LLM arenas rarely share code or citations, yet both reduce to the same four primitives: the pair queue, blinding and order, the vote record, and the standings. Comparing tools like BeaqleJS, listen-and-rate, and orq-arena reveals where the two worlds diverge. LLM harnesses judge pairs in both orders to discard inconsistent votes and control for response length, much like...
Why AI Leaderboards Dropped Elo for Bradley-Terry 14.11.2026 28:46
Turning pairwise votes into a reliable ranking requires more than counting wins. Replaying 135,632 real arena votes shows online Elo mostly remembers the last 22 games, shifting one model by 32 places compared to a batch Bradley-Terry fit. This episode unpacks the math behind Bradley-Terry, Elo, and Microsoft's TrueSkill, using a 120,000-vote Indic TTS study and LMArena's switch as real-world benc...
Why Comparing Two Things Beats a Rating Scale 14.11.2026 25:22
Humans are terrible at holding an absolute rating scale steady, which is why comparing two options side by side almost always wins. From psychophysics and 2AFC to ITU-T P.800 audio testing, Thurstone's 1927 law, and the Bradley-Terry model behind Chatbot Arena, four different fields invented the same trick under different names. Dropping a five-point scale for a one-bit choice cancels out rater fa...
Why Text-to-Speech and Speech-to-Text Aren't Mirrors 14.11.2026 27:33
Text-to-speech and speech-to-text look like mirror images, but the models underneath solve opposite mathematical problems. Synthesis expands a short string of characters into tens of thousands of continuous audio samples, inventing phoneme durations along the way. Recognition collapses a noisy waveform onto discrete words, discarding everything the ear doesn't need to distinguish meaning. This bre...
Chatterbox Retires as Open TTS Splits in Two 14.11.2026 23:21
Resemble AI has quietly retired Chatterbox, the permissive open voice model powering this very podcast, pivoting instead to deepfake detection. Meanwhile, the open text-to-speech landscape is splitting into two extremes. At the heavy end, Fish Audio's 35-billion-parameter Drama 3 and Boson's Higgs TTS 3 bring massive scale alongside restrictive community licenses. At the tiny end, nearly two-year-...
How One-Way Data Diodes Protect Power, Water, and Rail 14.11.2026 20:48
A data diode enforces one-way network traffic with physics: a fibre link with a laser on one end and no laser on the other. Originally a classified defence tool, these unidirectional gateways are now mandated in nuclear plants, used by municipal water utilities and rail operators, and required by France's ANSSI where firewalls are forbidden. This episode unpacks how protocol breaks keep TCP workin...
Adversaries vs. Auditors in Hardened Linux 13.11.2026 26:59
Half of what gets sold as hardened Linux isn't built to stop attackers—it's built to satisfy auditors with DISA STIG and CIS compliance checklists. Part one of this two-part look at hardened Linux separates compliance-focused builds from adversary-hardened systems like Qubes, Tails, Whonix, and Kicksecure. Each makes a distinct architectural bet: Xen-based compartmentalisation, amnesic USB session...
How Air-Gapped Linux Systems Actually Get Updated 13.11.2026 27:15
Almost nobody runs a special air-gapped operating system; instead, enterprises deploy mainstream Linux like Ubuntu with Landscape or Red Hat with Satellite in disconnected mode. The real engineering challenge is moving updates across the gap without breaking it. Because Landscape cannot generate diffs, keeping offline Ubuntu machines patched means physically carrying hundreds of gigabytes of repos...
How SecureDrop and GlobaLeaks Run in Practice 13.11.2026 26:19
Whistleblowing systems rarely fail on cryptography—they fail operationally. An audit of institution-grade tooling reveals that the real divide isn't open-source versus commercial, but supplied-and-supported versus self-deployed. The episode examines SecureDrop's two-server architecture and its transition to isolated Qubes OS workstations, alongside GlobaLeaks's dual Tor and HTTPS model for public...
What It Actually Costs to Run an Anonymous Tip Line 13.11.2026 22:11
Running an anonymous tip line without institutional backing comes down to custody, maintenance, and whether sources can even reach you. A real-world ledger compares SecureDrop, GlobaLeaks, Hush Line, OnionShare, Signal usernames, and raw Tor onion services—from SecureDrop's multi-thousand-dollar hardware list to GlobaLeaks' modest specs and Hush Line's hosted-first approach. It also examines why T...
How Survivor Charities Build Tech for Hostile Homes 13.11.2026 30:52
Domestic-violence charities face an inverted threat model: the attacker often shares a home and physical device with the person seeking help. This episode explores how small, underfunded organisations build safe digital front doors—from Women's Aid's cookie-free Live Chat and the US National Domestic Violence Hotline's disguised SMS billing to Tech Safety Canada's toolkit for avoiding third-party...
Testing Whether AI Rewrites Can Hide Your Identity 13.11.2026 29:52
Authorship obfuscation research shows large language models cannot guarantee anonymity when rewriting personal stories. While a FAccT conference whistleblower tool cut author attribution from 98.8% to 31.2%, it also erased a quarter of the story's meaning. Meanwhile, a plain Mistral prompt outperformed the specialized paraphraser DIPPER, and RANLP researchers found that LLM rewrites work well for...
When Changing Passwords Signals an Abuser 13.11.2026 20:39
Most technology-facilitated abuse does not rely on secret spyware, but on shared cloud accounts, Find My, car apps, and Bluetooth trackers set up while a relationship was fine. Because changing a password or removing a tracker triggers loss-of-access detection, standard cybersecurity advice can escalate danger. Drawing on guidance from Faraday, WomensLaw, TechSafety, and Tech Safety Canada, this e...
When Encryption Friction Nearly Stopped the Snowden Leak 13.11.2026 27:42
Edward Snowden's first anonymous approach to Glenn Greenwald stalled for months because Greenwald wouldn't install PGP, even after watching a 12-minute video tutorial. That friction exposes a deeper flaw in whistleblower tools: a 1999 study found only two of twelve experienced email users could correctly encrypt a message, and the 2020 paper 'Snitches Get Stitches' argues most leakers face small w...
AI for Mental Health and Our Encrypted Tip Line 13.11.2026 18:24
When therapy is out of reach, carefully built AI tools can serve as an invaluable mental health ally rather than an empty room. Following nine listener episodes on toxic and narcissistic family systems, this wrapper explains how language models can anonymise sensitive stories and how our new Hush Line tip line works. Built by the nonprofit Science and Design, it uses browser-based PGP encryption,...
Podcasturi asemănătoare
Replaio nu este editorul podcasturilor; numele emisiunilor, coperțile și materialul audio aparțin autorilor lor și sunt distribuite prin fluxuri RSS publice