Every big local-AI headline, run down to the source. Red means the claim is
busted, green means it checks out, amber means it's more complicated than the headline let on.
Each verdict opens to its receipts, and every one links to the video where we walk through it.
Read the sources yourself — that's the point.
Real progress, wildly oversold. GPT-5.6 "Sol" set a genuine ARC-AGI-3 record and was the first model to fully win ONE game — but that's 7.8% vs an untrained human's 100%.
Receipts4 sources
GPT-5.6 Sol: 7.78% semi-private / 13.33% public, new SOTA, first to fully win a game (FT09, 152 actions vs 208 human): arcprize.org/results/openai-gpt-5-6-sol + @arcprize
Human baseline 100%, median 7.4 min untrained: ARC-AGI-3 Technical Report (arcprize.org)
March 2026 best was Gemini 3.1 Pro at 0.37%: officechai.com
~$20,000 per max-reasoning eval: thealgorithmicbridge.com
“The AI bubble is bursting — it all crashes next week.”
MIXED
The froth in AI valuations and capex is real — but "it all crashes next week" is an unprovable guess, and the part that survives the frenzy is the local, open AI you run yourself.
“It tops SWE-Bench, so it's the best AI coder money can buy.”
BUSTED
The leaderboard is real, but OpenAI's own audit found roughly a third of SWE-Bench Pro's tasks are broken — so a headline coding score can't reliably prove which model is actually the best engineer.
Receipts6 sources
OpenAI "Separating signal from noise in coding evaluations" (Jul 8 2026; ~30% of SWE-Bench Pro broken, recommendation retracted): openai.com/index/separating-signal-from-noise-coding-evaluations
“Hand over your health data to Samsung's AI, or it wipes everything.”
MIXED
The scary dialog was real; the "hands over your data or we wipe everything" mass-deletion was NOT.
Receipts6 sources
Dialog + consent wording: 9to5Google (Jul 13)
Samsung's clarification: 9to5Google (Jul 15)
Four data categories: Android Authority
"Decline doesn't delete your data": Android Police / SamMobile
Export/erase how-to: Samsung support ANS10001379
Unpacked Jul 22: SamMobile
Watch the check →Did Samsung REALLY Threaten to Delete Your Health Data? · 1:21
“A free open model beats GPT — just download it and run frontier AI at home.”
MIXED
The open weights really are free — but GLM-5.2 needs 200GB+ to even load, so for almost everyone it's a rented API, not something your gaming PC can run.
Receipts5 sources
744B params + benchmark wins (SWE-bench Pro 62.1 vs 58.6): VentureBeat, apidog
Memory to load: 2-bit ~245GB (avenchat), INT4 372GB (Spheron)
RTX 5090 = 32GB current flagship: nvidia.com
API pricing (~$1.40/M in, $4.40/M out): apidog/OpenRouter
