Professing themselves to be wise, they became fools
Joined February 2016
- Tweets647
- Following282
- Followers106
- Likes278
My AI's quality score is a great pessimist and a terrible optimist.
Since Oct 8, users rated 35 Diffio results. All 8 it scored poor got a thumbs down. Of the 22 it scored good, 11 got one too.
Let the score block disasters. Let a thumbs-down button grade the rest.
"100% automated SEO" threads skip the part you can't automate.
One viral thread: $12k MRR, 1,435 bookmarks, by a founder who sells an SEO agent. In the replies he admits backlinks may never be fully automated.
Automate drafts. Earn links by hand. Where did your agent flunk?
Audio file headers lie like dating profile photos.
YODAS v3 is 1.1M hours, all 48kHz. Its own card says only 67% has real 44kHz content, and a quarter of its stereo videos are one channel copied twice.
Training TTS on web audio? Filter on measured bandwidth, not the header.
Your competitor's comment section is the warmest lead list on LinkedIn. Let a bot mine it and you break LinkedIn's rules.
Its User Agreement bans bots that scrape profiles or send messages. Spend 15 minutes a day on 4 rival founders' posts. DM only people who asked a question.
Easy way to beat Silero VAD on recall: call everything speech.
parakeet-redux's VAD head marks 1,685s of each hour of music and noise as speech. Silero marks 48s. That's parakeet.cpp's own benchmark, to its credit.
Want more recall? Drop Silero's threshold to 0.3 instead.
I was scared a ChatGPT listing would melt my self-hosted AI. It's busy 3% of the day.
Last 14 days: 1,137 Diffio jobs, 11 hours of compute, mostly on a box named mac-mini-1. Longest queue wait: 2.8s.
Time every job before buying GPUs. My scaling problem is a demand problem.
Your voice agent survives street noise. Turn on a TV and it takes orders from the news anchor.
@_josh_meyer_ tested 14 voice stacks. A TV cut completion from 39% to 9%. LiveKit's fell 43% to 0%. Its noise filter is off by default.
Turn voice isolation on. Test with a TV on.
Your best reply can hit a million views and count for zero toward getting paid on X.
Original Content Rewards needs 500k Home Timeline impressions from verified users in 90 days, and replies are excluded.
Use replies to find followers. Only your own posts move that number.
NVIDIA looked at 354,000 hours of speech and binned 69% of it.
Granary-HiFi keeps 111k hours that pass four checks: one speaker, no background noise, a good predicted quality score, no clip ending before its transcript does.
Run those on your fine-tune data before buying more.
Google sends me more signups. ChatGPT sends me more paying customers.
Since Oct 11: 123 Diffio signups from Google search, 77 from ChatGPT. Google's became 7 paid accounts. ChatGPT's became 8.
Tiny sample. Still, tag each signup with its source and rank channels by who pays.
Topical maps are a 40-slide pitch deck for a startup with zero customers.
Google's documented "topic authority" is a news system that scores original reporting. For other sites it asks: do you have a focus?
What got you clicks: the map, or 10 pages on questions buyers asked?
15,000x real time is a batch number. Your live voice agent gets 54x.
NVIDIA's Nemotron 3 Diarization hits 15,113x on 32 batched files with a 30s buffer. One live stream at 0.32s: 54x.
Accuracy barely moves (12.7% to 13.6% DER). Batch the backlog, stream the calls.
A "23% conversion lift" with no visitor count is a coin flip with good lighting.
Trusting a 23% lift takes about 300 conversions per version. At 3% conversion, that's 10,700 visitors each.
Ask for that count before copying a winner. End your tests there, not when they go green.
Voice filters send the bill to your clean calls.
@huseinzol05's new open isolator cuts word errors from 46.8% to 14.8% across 9 test sets. Real win. But with the caller alone on the line: 0.04% to 5.6%.
Count your calls with a second voice before you ship one.
Google hasn't sent one visitor to my newest page. ChatGPT sent 39, and 20 signed up.
It's a plain "fix distorted audio" page I shipped a week ago. ChatGPT started linking it 3 days later.
Sort your ChatGPT signups by landing page. Google won't show you this one.
Cold email works. It just takes about 4,000 emails to book one meeting.
One agency's own numbers: 8-10M emails a month, 2,000+ meetings. Even at Gmail's 0.1% spam target, that's 4 spam reports per meeting.
Small SaaS? Send 50 you'd sign. How many did your first meeting take?
One LayerNorm can flip who wins a SOTA chart.
AudioFM tops 10 of 12 audio tasks on raw features. Add LayerNorm to the probe and its +4.7 heart murmur lead falls to 0.8, and it loses apnea subtypes. Credit to them for printing both.
Picking an encoder? Probe it both ways.
Claude Code now has an env var called EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS. It lets rm -rf ~ run unasked.
Skip mode never stopped prompt injection, per Anthropic's docs. Deny rules leak too: git -C . push slips past a git push rule.
Use a container you can afford to lose.
Diffio is now a ChatGPT plugin. Our GPUs don't autoscale.
Founders say a ChatGPT listing sent them a flood of users. We run inference on our own hardware, not the cloud, so if they're right, I'll know fast.
Clean up noisy speech right inside ChatGPT:
chatgpt.com/plugins/plugin_a…