Finally! A Local AI Breakthrough! So Much Faster!
Credibility score: 52/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "Finally! A Local AI Breakthrough! So Much Faster!" and rated it 52/100 for credibility (a BS score of 48/100 — mixed credibility), on 2026-09-17. Its weakest claim — "Dismisses benchmarks, claiming 'real use' is the only valid metric." — scored 20/100 and was flagged as false dilemma. 30 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Claims analyzed
Declares a "major breakthrough" in local AI — setting the stage with hype. — Loaded Language (45/100)
A 'major breakthrough' is a hell of a claim to open with — no details, just pure, unadulterated hype. 😈
Claims 'stars aligned' and local AI is 'ready for prime time' — vague, confident pronouncements. — Confidence Mismatch (45/100)
The 'stars aligned' and 'overnight' changes? That's not a technical explanation, that's a fairy tale for mortals. 💀
Credits 'community tweaks' for AI speed improvements — vague attribution. — Anonymous Authority (45/100)
Credits 'community tweaks' without naming a single one. 'Others' is the weakest authority in my book. 🔥
Claims his AI machine is '10 times faster' due to community tweaks. — Confidence Mismatch (45/100)
Ten times faster? That's a hell of a jump from 'community tweaks.' Sounds like a number pulled from thin air. 💀
Declares local AI 'actually useful' in '1 week' due to new models and speed. — Confidence Mismatch (45/100)
One week, and suddenly local AI is 'actually useful'? That's a quick turnaround for a 'breakthrough.' 🙄
Dismisses benchmarks, claiming 'real use' is the only valid metric. — False Dilemma (20/100)
He's setting up a false choice: either benchmarks OR 'real use.' You can do both, mortal. One informs the other. 😈
Reiterates dismissal of benchmarks, defining 'real work' as the sole measure of success. — Confidence Mismatch (45/100)
Still no benchmarks, just 'real work' that 'fails or is incompetent.' That's a subjective metric, not a standard. 🔥
Sources: Why Beating the Benchmark Isn’t Always the Right Measure of Success - CY Actuaries, Benchmarking Basics: Everything You Need to Know About Job Benchmarking - TTI Success Insights, Escaping the Benchmark Trap: A Guide for Smarter Investing
Boasts about the 'sophistication' of his AI use. — Just Vibes (50/100)
Oh, 'amazed,' are we? That's not a claim, that's just main character energy. 💅
Sources: Donald Trump Refuses to Say What He Uses AI for While Talking to Reporter: Watch, Trump claims HE'S the only guardrail that AI needs thanks to his IQ, What Trump’s Bizarre Posts Are Teaching Us About AI - The Atlantic
Claims his Brax.me site's tech support is handled by a '100% local AI' agent. — No Frame (75/100)
Finally, a specific example. He's claiming his tech support is fully local AI. Let's see if he actually shows it. 😈
Mentions his 'Brax.me site' — a subtle brand drop. — Plain Sales Pitch (45/100)
He just slips in his own site name like it's common knowledge. A little self-promotion, I see you. 😈
Pitches Ollama Pro subscription — presents it as a solution. — Sponsored (50/100)
Oh, a $20/month subscription? And he just *happens* to mention it. This isn't a recommendation, it's a soft sell. 💸
Declares a $4,000 AMD Strix Halo machine the 'current sweet spot' — dismisses other options. — Confidence Mismatch (45/100)
A $4,000 machine is the 'sweet spot' and anything cheaper is 'ineffective'? That's a bold claim without showing any actual benchmarks. 💀
Claims 96GB is 'plenty' but then says models aren't tuned for it. That's a contradiction, mortal. 💀 — Volume Game (45/100)
Says 96GB is 'plenty' for a model, then immediately adds that no one's making models for that size. So, is it plenty or not? Pick a lane. 🔥
Declares GPT-OSFS-120B the 'most effective' for reasoning, despite its age. Bold claim, zero comparison data. 😈 — Confidence Mismatch (45/100)
Calls a year-old model 'the most effective' for reasoning without showing any actual benchmarks against newer, smaller models. That's just a feeling, not a fact. 💀
Claims 'long-term viability' for a model by adding web search and data. That's not viability, that's a patch. 😈 — Confidence Mismatch (45/100)
Says adding web search and local data gives a year-old model 'long-term viability.' That's like putting new tires on a rusty car and calling it future-proof. 💀
One of two models 'can coexist' with the main one. That's not a breakthrough, that's just fitting. 😈 — No Frame (75/100)
Just states that one of the tested models can run alongside the main one. It's a straightforward observation about resource management. ✅
Calls 'usable models' an 'alignment of the stars' and 'brand new.' Overdramatic for basic functionality. 💀 — Loaded Language (45/100)
Describes simply 'having usable models' as an 'alignment of the stars.' That's not a cosmic event, mortal, it's just progress. 🔥
Defining 'prefill' as input tokens per second — just setting the stage. — No Frame (75/100)
Just defining terms, nothing tricky here. A rare moment of clarity. 🔥
Declaring input as the 'major bottleneck' — a confident assertion without immediate data. — Confidence Mismatch (45/100)
He just declared 'input' the 'major bottleneck' like it's obvious. Where's the proof, mortal? 💀
Exaggerating local AI interaction time to 'closer to an hour' — dramatic comparison. — Emotional Button (45/100)
He jumped from '4 to 5 minutes' to 'closer to an hour' for local AI. That's not a comparison, that's a horror story. 😈
Claims a massive speed increase for local AI, citing specific TPS numbers. No Frame. — No Frame (75/100)
He's throwing out specific numbers for speed increases, 688 prefill TPS and 53 output TPS. That's a direct claim, not a trick. 🔥
Translates raw speed into practical time savings, declaring local AI 'useful.' No Frame. — No Frame (75/100)
He's converting the TPS numbers into real-world time savings, making the claim tangible. He's saying it's 'finally' useful. 😈
Details his specific hardware and model setup, including memory usage. No Frame. — No Frame (75/100)
He's laying out his exact setup: Strix Halo, two specific models, and their memory footprint. That's just information, not a trick. 😈
Claims GPT-OSS size model fits Strix Halo, largest for local AI without strain. — No Frame (75/100)
He's just setting the stage for what's possible with current hardware. No tricks, just facts. 🔥
Introduces Qwen-3.8 Flash, same size as GPT-OSS-120B, but admits it 'didn't work'. — No Frame (75/100)
He's just laying out the options and his own experience. Refreshing honesty, for a mortal. 😈
Declares MoE models are the ONLY choice for local AI if speed is 'paramount'. — False Dilemma (20/100)
He says 'no choice' like there aren't other factors. Speed isn't the only god, mortal. 💀
Claims MoE models are 'often five times faster' than dense models. — Confidence Mismatch (45/100)
He says 'often five times faster' like it's a universal law. 'Often' isn't 'always,' mortal. 🚩
Claims Qwen 3.6-35B has 1071 TPS prefill, 50% faster than GPTOSS. — No Frame (75/100)
He's giving a specific metric and a direct comparison. This is how you back up a claim, mortals. 🔥
Sets up a straw man about prefill times — then knocks it down. — Straw Man (20/100)
He's telling you what 'you might think' about prefill times, then immediately contradicting it. Classic setup. 😈
Claims local AI prefill is 'pretty fast' after first load — almost cloud speed. — Confidence Mismatch (45/100)
He says 'pretty fast' and 'almost as fast as cloud' — but '1 or 2 seconds' isn't 'almost' cloud speed for prefill. That's a big gap. 💀
See the full analysis with sources and timestamps →