Google Just Found a Loophole in AI Hardware Limitations
Credibility score: 55/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "Google Just Found a Loophole in AI Hardware Limitations" and rated it 55/100 for credibility (a BS score of 45/100 — mixed credibility), on 2026-06-09. Its weakest claim — "QAT models keep near-full intelligence with far fewer parameters" — scored 40/100 and was flagged as dubious. 12 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 12 claims analyzed: 0 scored under 40, 9 between 40 and 69, and 3 at 70 or above.
Claims analyzed
12B Gemma bridges both size and capability gap — OK (60/100)
At 2:30
Sounds plausible but no benchmarks shown yet — just the speaker's read.
Why this score: Size bridge is real — transcript confirms 12B sits between E2B/E4B and 31B. Capability claim is softer:
- Speaker says it improves "capability" without numbers
- No specific benchmark comparison given in segment
- Web results only mention the model exists, no performance data yet
Original quote: “the 12B model was interesting because it is bridging the gap in not only size in terms of offering for this family, but also capability”
12B unified beats E models "without thinking" — Dubious (45/100)
At 4:30
"Significantly better than thee without thinking" — the phrasing is doing the work here; charts aren't quoted.
Why this score: Vague comparison.
- Speaker claims 12B outperforms E2B/E4B "without thinking"
- No specific metric or delta provided
- "Without thinking" appears to be casual shorthand, not a defined test
Original quote: “And an interesting detail is that the 12B unified is scoring significantly better than thee without thinking”
Qwen 3.5 9B beats Gemma 4 12B on published benchmarks — Dubious (45/100)
At 6:30
Mentions an unnamed benchmark mashup — zero link, zero numbers.
Why this score: Anonymous benchmark red flag.
- Speaker says Qwen wins but never names the test or the post
- Community comments echo the same vibe (Qwen stronger on coding/agentic tasks) but still no source
- No evidence provided = low confidence
Original quote: “someone published this kind of like you know benchmark mashing uh between the 12B unified and the Quinn 3 and a half 9B and Quinn does do better on these benchmarks”
QAT isn't a new Google invention — Solid (75/100)
At 8:30
QAT predates Google frontier releases by years — checks out.
Why this score: Accurate history.
- QAT papers date back well before current frontier labs
- Speaker correctly notes Google didn't originate the technique
Original quote: “training QAT. Now as I've mentioned this is not like some novel concept Google has invented.”
This is the first frontier lab using QAT at scale — Dubious (45/100)
At 8:37
Meta already applied QAT to original Llama models — not first.
Why this score: Timeline mismatch.
- Speaker later cites Meta's early Llama QAT work
- Contradicts the 'first frontier lab' framing
Original quote: “It just hasn't been done with a big frontier model lab release like this. And QAT kind of changes the mental math that it takes to run these models.”
Meta used QAT on the first Llama models — Solid (80/100)
At 8:50
Meta did post-training QAT on original Llama — speaker's own source backs it.
Why this score: Consistent with record.
- Early Llama papers include quantization-aware fine-tuning steps
- Matches known timeline before Llama 2
Original quote: “one of the earliest papers I can find is when the meta meta researchers were actually doing this with the llama series of models, not even llama 2, just the original llama models”
BitNet is a subset of QAT, not the same thing — Opinion (50/100)
At 10:30
Calling BitNet 'maybe a subset' of QAT is his read, not a settled fact.
Why this score: Personal framing.
- Speaker says 'hazard to say' and 'maybe' — already signaling this is interpretation
- No sources or definitions provided to anchor the subset claim
- Technical overlap exists but relationship between BitNet and QAT remains debated in the field
Original quote: “I would hazard to say no. Uh I think it is maybe a subset of QAT because it is very similar for sure.”
BitNet trains weights to only 1, 0, and -1 — Solid (75/100)
At 10:37
That matches the published BitNet paper description of ternary weights.
Why this score: Accurate description.
- BitNet b1.58 paper explicitly uses weights restricted to {-1, 0, +1}
- Speaker correctly identifies the core constraint that defines the approach
- Minor rounding in wording but substance holds
Original quote: “Whereas you basically train a model under the assumption that hey your weights will be you know 1 0 and negative 1.”
Prism ML models use post-training quantization like QAT — Dubious (45/100)
At 10:48
No public details confirm Prism ML's exact training method — this is speculation.
Why this score: Unverifiable comparison.
- Prism ML has released limited technical information on their training pipeline
- Speaker frames it as 'might be' and 'almost' — hedging on missing evidence
- No citations or papers referenced to support the parallel
Original quote: “this is something that actually might be more similar to how these models from prism ML have been trained where it's like almost quantization aware right it's an additional step after the model's been trained”
Google's 12B Gemma QAT is different from 1-bit/ternary models — OK (60/100)
At 10:58
True that Google's approach isn't full 1-bit, but he gives zero specifics on how it differs.
Why this score: Partial distinction.
- Gemma 4 12B QAT is indeed not a pure ternary or 1-bit model
- Speaker correctly notes the difference but provides no technical breakdown
- Distinction is real but remains at the level of assertion without evidence shown
Original quote: “one bit models, turnary models maybe are similar to QAT. They're definitely not the same as what Google has done with the 12B.”
QAT models keep near-full intelligence with far fewer parameters — Dubious (40/100)
At 11:10
'Way, way, way less parameters' while staying 'just as smart' is the usual QAT marketing stretch.
Why this score: Overstated tradeoff.
- Real QAT results show measurable quality drop even at 4-bit
- 'Way less parameters' phrasing conflates bit-width reduction with actual parameter count
- 'Hopefully just as smart' is the speaker's own hedge on the claim
Original quote: “it allows you to run a model that's just as smart, hopefully just as smart, uh, with very little loss compared to its non QAT kind of sibling, but you run it on way, way, way less parameters.”
This is similar to BitNet / 1-bit / ternary models that already showed promise — Dubious (45/100)
At 16:30
Calls it "similar but maybe not the same" — the connection stays vague.
Why this score: Loose comparison.
- Mentions BitNet, 1-bit, and ternary models as prior examples
- No technical overlap details given
- Web search shows no confirmed link between Gemma QAT and those architectures
Original quote: “That's really cool. the fact that this is also similar but maybe not the same as bitnet or the uh one bit models or the turninary models which we've seen promise with already where we have I think it was flux 2 clin image models from prisml the bonsai image gen which I did a video I'll link here as…”
See the full analysis with timestamps →