Professing themselves to be wise, they became fools
Joined February 2016
- Tweets264
- Following283
- Followers95
- Likes261
The hot new SEO page grader is an AI that agrees with one consultant and three other AIs 82% of the time, if you count near misses.
Fine as a cheap editor. Before you trust any AI grader, ask for exact-match agreement and what always answering "Medium" would score.
Pulling the dialogue out of a movie used to need a GPU and a grudge. Now it's a 1M-parameter model on a CPU.
DBS-TIGER gets speech out at 15.9 dB SI-SDR. Music and effects trail by 8 to 9 dB, and its speech training data is English audiobooks, so "split ANY soundtrack" is generous.
Dialogue is the easy stem now. What would you build if every video came with a free dialogue track?
Turns out you can overthink your way out of a right answer, even as a transformer.
@pixeli99's InfiLoop paper calls it latent overthinking. Later loops overwrite deductions earlier loops got right. Their fix learns how much of each update to accept. A 7M-param model kept improving past 20,000 steps on Sudoku-Extreme, hitting 97.9%.
Agent loops do the same thing. Append each pass to a log, and only overwrite what a check says is broken.
bytes/4 is a fine tokenizer. I tested the one in @mattpocockuk's 5,000-token file cap against OpenAI's tokenizer on 600 Python and JS files.
It missed 15 of 141 files over the cap. 13 were data: encoding tables, number-heavy tests, an emoji regex. Skip data paths, keep bytes/4.
7/ I supplied screenshots, a screen recording, organization details, timestamps and browser info.
What’s missing now is an account-level answer: why this is blocked, whether it’s intentional, and what path is actually supported.
Another help-center link is not a diagnosis.
8/ @OpenAIDevs @OpenAI: is the blocked upgrade an account defect or a grant restriction? What supported path lets me increase production limits without forfeiting the grant? Who owns the answer?
World-changing products. Customer support stuck on repeat.
Please break the loop.
Force-indexing backlinks is the SEO tip of the week. A few cents a link, on Google in minutes.
I checked. IndexCheckr tracked 33,930 pages sent through five paid indexers. 29% got indexed.
A ping can bring Googlebot over. It can't make Google want the page.
Tested @_josh_meyer_'s "speech, not noise" point at toy scale: 40 TTS lines, a second voice as the TV, Whisper base. At 10 dB, matched speech-shaped noise: 4% WER. Talker: 38%, mostly TV words before/after the caller. Crop to the caller's own span: 5%. Turn edges are the leak.
Tested @_josh_meyer_'s "speech, not noise" point at toy scale: 40 TTS lines, a second voice as the TV, Whisper base. At 10 dB, matched speech-shaped noise: 4% WER. Talker: 38%, mostly TV words before/after the caller. Crop to the caller's own span: 5%. Turn edges are the leak.
Put classic denoisers in front of Whisper base (40 TTS lines + pink noise, self-run). At 0 dB, untouched audio: 7.8% WER. Spectral subtraction 10.9%, Wiener 11.6%, gating 14.7%, median filter 25.9%. By 10 dB they're within a few words. If ASR is the listener, check WER first.
M17 is trying a neural synth on Codec2 3200 frames, same bitstream. I pushed 40 TTS lines through c2enc/c2dec: 8 kHz alone scored UTMOS 4.22, WER 4.1%. Codec2 3200: 2.96, 5.0%. The words survive and naturalness is what drops, so a receiver-side synth targets the right gap.
Cartesia's per-entry case for pronunciation entries (IT vs it) fixes a real collision. I ran 24 sentences with IT/US/WHO/AM/OR/SAT through flite and espeak-ng:
No lexicon: flite says 12/12 acronyms as words
Case-blind alias: spells 12/12 plain words
Per-entry case: 0 wrong
