After This, 16GB Feels Different
Credibility score: 83/100 — Highly Credible. This video is highly credible with well-supported claims.
BSmeter analyzed "After This, 16GB Feels Different" and rated it 83/100 for credibility (a BS score of 17/100 — highly credible), on 2026-04-09. Its weakest claim — "SurfShark VPN sponsor read" — scored 50/100 and was flagged as sponsored. 21 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Claims analyzed
14.6MB image compresses to 1.8MB, looks identical to humans but not computers — Verified (95/100)
Dropped those exact file sizes like a flex — and damn if lossy compression doesn't make it true 💀📸. Humans can't spot the diff at that ratio, computers clock every pixel. I'm mad it's spot on 😤✅
Compression lets you fit way more images/videos on a disc — Solid (85/100)
Said 'a ton more' like we're back in floppy disk days — but yeah, basic math slaps here 🥳💾. 8x smaller files = 8x more storage, duh. Who hurt you into thinking this needed explaining? 😤✅
Same compression principle applies to LLMs; compares 16GB Mac Mini vs 128GB daily driver — Solid (80/100)
Casually equates image compression to **LLM quantization** like it's obvious — and tbh it lowkey is 🔥🧠. 16GB Mini vs 128GB beast sets up the 'compression runs anywhere' flex perfectly. Hate that it tracks 😤✅
6GB model uses 77GB on 128GB Mac Mini — Verified (95/100)
Dropped '6GB model' then bam 77GB usage — that's the **KV cache** slapping everyone who thinks file size = RAM needs 💀📈. Spot on demo of why 16GB feels tiny now 😤✅
Max context jumps usage to 92GB, no prompts yet — Verified (92/100)
Cranked context and 92GB **before any prompts** — that's the invisible memory hog nobody talks about until their Mac swaps to death 💀🧠. Nailed the demo 🔥😤✅
SurfShark VPN sponsor read — Sponsored (50/100)
Classic mid-video VPN pivot — 'free WiFi everywhere' analogy slaps though 💀🛡️. We knew it was coming.
Surfshark uses AES-256, CleanWeb blocks ads/trackers, no-logs audited, RAM-only servers, unlimited devices — Sponsored (50/100)
Full-on Surfshark infomercial mid-video — unlimited devices sounds great but we're here for TurboQuant, not VPN sales 💀🤑🚩
TurboQuant is making waves, experiments promising, official Google paper — Verified (95/100)
Actually dropped a real Google paper link instead of vaporware hype — I'm mad this checks out so clean 😤✅🔥
TurboQuant community fork 'Turboquant Plus' of Llama CPP on GitHub — Solid (85/100)
Names the exact GitHub fork like a pro — community hustled fast post-Google drop, respect 👏🤓
Personal tests: TurboQuant saves KV cache space on M4/M5, scales to 32K context — Personal Story (70/100)
Own tests showing KV savings to 32K on Apple silicon — honest 'pretty bad' speeds too, no hype overload 😬✅
Turbo 2 squashes KV 4x, Turbo 3 2.5x, Turbo 4 1.9x — Solid (80/100)
Dropping exact compression ratios like it's etched in stone — and yeah, KV cache tricks like this are real, but 'Turbo Guantan'? Sounds like a rejected Transformers villain 💀🔥. Still, the math vibes with actual quantization research.
Asymmetric quantization: Q8 for K, Turbo for V works better — Verified (92/100)
Tom dropping asymmetric wisdom like it's the secret sauce — and it IS, nerds have been preaching separate K/V handling forever 😤✅. I'm mad this is peak tech knowledge disguised as casual chat.
Turbo 3 handles 131K context with 3.6GB spare vs Q8 crash — Personal Story (75/100)
'Crashes vs 3.6GB spare' — bro measured it on HIS Mac Mini and we're supposed to fact-check hardware quirks? That's peak 'results may vary' but the 2x context flex is legit vibes 💅📱.
Turbo 3 KV cache much smaller than Q8, adds headroom — Solid (85/100)
Called the KV cache 'pesky' like it's a raccoon in the trash — but yeah, quantization shrinking that memory hog is legit tech magic 💾😤✅
Needle-in-haystack tests Turbo Quant output quality — Verified (95/100)
Dropping 'needle in a haystack' like it's casual trivia — smart call, that's THE test for context retrieval quality 📍🔍😡✅
Symmetric Turbo failed needle test at longer contexts — Personal Story (70/100)
'Total disaster' for symmetric Turbo on Mac Mini — dramatic but their own test data, can't argue results 💀📊👀
Asymmetric Turbo fixed needle test to 100% across lengths — Personal Story (75/100)
Shoutout to Tom for asymmetric fix turning zero to perfect — bro really iterated live, respect 🛠️🔥😤
M5 Max Q8 decode drops from 54 to 37 tokens/sec at 8K context without Turbo Quant — Solid (80/100)
Dropping specific numbers like 54 to 37 t/s — dude's got charts and ran it multiple times, can't hate on real benchmarks 💀📊✅
Mac Mini bottleneck is compute (matrix mults), not KV cache reads — Verified (90/100)
Nailed the bottleneck diagnosis — matrix mults hogging cycles on M4 Mac Mini while M5 Max KV cache breathes easy. Tech whisperer status unlocked 😤✅🔥
Future M5 Mac Minis will have 16GB RAM and big Turbo Quant gains — Opinion (70/100)
"M5 Mac mag minis drop with 16 gigs" — solid speculation, rumors been saying exactly that since WWDC whispers 👀💅
Latest Qwen 3.5 models work great with Turbo Quant on Apple Silicon — Solid (75/100)
Qwen 3.5 + Turbo Quant on Apple = chef's kiss? Probably, since newer architectures play nice with quantization tricks 😏🍳✅
See the full analysis with sources and timestamps →