RIP Claude Fable, open-source AI unleashed, full body avatars, new Google models, new TTS: AI NEWS

Scale 2 beats Cling 3 at motion transfer, handles multi-character scenes and camera movement — Opinion (50/100)

At 0:30

Subjective quality ranking with side-by-side demos — no external benchmarks cited.

Why this score: Pure opinion based on visual demo. - Speaker shows examples but provides no third-party tests or metrics. - "Best open-source" and "on par with Cling 3" are personal judgments, not measured claims.

Original quote: “First up, we have probably the best open-source AI for transferring motion from one video onto another video. It's called Scale 2... note that this new Scale 2 is just a lot better. I'd say the quality is even on par with the closed source Cling 3... Scale 2 is even able to capture the camera…”

Actionable world model handles hands, humans, earphones, and Unitere dog robot — Unverifiable (50/100)

At 2:30

Demos sound specific but no independent confirmation of the Unitere dog example — taking his word for it.

Why this score: Demo claims are too new to verify. - Project described as recently released - No matching results for "Unitere dog robot" in current search - Other object types (hand, human, deformable items) are standard in this research area

Original quote: “And what makes this impressive is the range of objects it can handle. The demos show articulated objects like a hand or an entire human moving to things like deformable earphones or even the Unitere dog robot.”

Says Oscar's code and run instructions are released on the page — Unverifiable (50/100)

At 4:30

No link or repo shown on screen — can't verify the code is actually public.

Why this score: No evidence provided in the segment. - Speaker claims code is available but doesn't display the repo or instructions - Apache 2 license mentioned but can't confirm without the actual page

Original quote: “at the top of the page, they have released the code to this. So, if you scroll down here, it contains all the instructions on how to run this on your computer.”

Sources: Documenting Code · Oscar.jl

Gemini detects over 70 languages automatically — OK (60/100)

At 6:32

Sounds plausible but no source given — just tossed out as fact.

Why this score: No citation provided. - Google has claimed high language coverage before, but exact number needs verification - Transcript gives no link or study backing the 70 figure

Original quote: “Note that here it can automatically detect over 70 languages.”

Translation keeps original intonation, pacing, pitch — Dubious (45/100)

At 6:38

Real-time translation usually flattens tone — hard to keep natural voice intact.

Why this score: Voice preservation is the weak point. - Current systems struggle to maintain speaker-specific prosody across languages - Demo clips sound clean but may be edited or cherry-picked

Original quote: “And it also tries to preserve the speaker's intonation, pacing, and pitch. And it also tries to keep the conversation natural.”

Gemini translates with only seconds of lag, not turn-by-turn — OK (55/100)

At 6:48

Continuous mode exists in research, but real-world lag is usually longer than 'a few seconds.'

Why this score: Latency claim is optimistic. - Live translation demos often have 3-8 second delays - No independent benchmarks shown for Gemini 3.5 specifically

Original quote: “Unlike turnbyturn translation systems that wait until someone finishes talking, Gemini 3.5 L translate generates continuously while staying only a few seconds behind the speaker.”

Feature already live in API and AI Studio for free — Solid (75/100)

At 6:58

Direct link to try it — rare they actually ship the demo tool same week.

Why this score: They gave a working link. - Google AI Studio is publicly accessible - Claim is testable right now rather than 'coming soon'

Original quote: “So that means fewer awkward pauses and smoother back and forth. The awesome thing is this feature is already available via their API as well as Google's AI studio which you can try for free.”

Google live translate now free for everyone on Android/iOS — OK (60/100)

At 8:30

Says it's rolled out to everyone — no source or confirmation date given.

Why this score: Unverified rollout claim. - Transcript states availability via Google Translate but gives no rollout date or official link - No corroborating search results confirm a recent free public launch - Could be accurate but lacks any supporting evidence here

Original quote: “So simply click on this link and you can try it here. And apparently this is available for everyone via Google Translate on Android and iOS.”

Google's diffusion Gemma is 4x faster than standard models — Dubious (45/100)

At 8:45

Google claims 4x faster — but that's their marketing number, not independent benchmarks.

Why this score: 4x speed claim is Google-sourced. - Speaker attributes the figure directly to Google with no third-party verification - No actual benchmark data or papers cited in the segment - Diffusion models historically struggle with text quality, so raw speed claims need scrutiny

Original quote: “Also this week, Google releases diffusion Gemma. This is their open-source model which is incredibly fast. Now, this is a bit different from regular language models which you may be used to. You see, most language models write text from left to right, one token or word at a time. But this diffusion…”

Diffusion Gemma now matches regular Gemma 4 on MMLU and GPQA — Unverifiable (50/100)

At 9:05

Says it matches Gemma 4 on benchmarks — but the actual numbers or chart aren't shown here.

Why this score: Benchmark parity claim lacks data. - Speaker references 'these benchmarks' and a light blue bar but shows no visible numbers - No MMLU or GPQA scores provided in transcript - On-screen chart mentioned but not described, so the claim can't be verified from audio alone

Original quote: “And you know the really cool thing is you can even get this to do pretty complex stuff like solving sudoku or generating 3D models as you can see from this example. Now historically these diffusion language models have performed pretty badly in terms of generating text reasoning and world…”

StreamForce hits 16.6 fps on single CPU — Unverifiable (50/100)

At 10:30

16.6 fps on one CPU sounds ambitious — no public benchmark yet to check.

Why this score: No independent verification. - Claim comes only from the project page - Code still unreleased - Real-world single-CPU performance for video gen is rarely this high

Original quote: “The page says that streamforce runs up to 16.6 frames pers on just a single CPU.”

Agents Last Exam benchmark ranks GPT 5.5 Codeex above Claude Fable 5 — Unverifiable (50/100)

At 12:30

No public results or leaderboard for 'Agents Last Exam' in the sources — just the speaker's word on the ranking.

Why this score: No independent verification found for this benchmark or its leaderboard. - Web results mention Claude Fable 5 being pulled offline by US government order, not model performance comparisons - No mention of 'Agents Last Exam,' 'GPT 5.5 Codeex,' or 'OpenClaw' rankings - Speaker's personal preference for GPT 5.5 Codeex noted but doesn't substitute for public data

Original quote: “have a new very important benchmark called agents last exam. You see this is basically a stress test for AI models doing real professional work not just answering random knowledge questions or solving tiny coding tasks. The idea is simple. If AI agents are supposed to help with actual jobs, then we…”

Arbor outperforms standard harnesses like cloud code or codecs across categories — Unverifiable (50/100)

At 14:30

Claims Arbor beats cloud code and codecs — no actual benchmark numbers shown on screen.

Why this score: No specific data provided. - Speaker says 'pretty significant' gains but never quotes the actual scores - Mentions categories (optimizer design, math reasoning) without linking to results - GitHub link offered but no numbers here to verify

Original quote: “And some of these gains are pretty significant and that's why this arbor system is so interesting.”

Kimmy released K2.7 Code model this week, top open-source lab — OK (60/100)

At 16:30

Name's off — it's Kimi, not Kimmy, and the model is actually Kimi K2.7.

Why this score: Name error on the lab. Key issues: - Real lab is Moonshot AI's Kimi, not "Kimmy" - Model name matches Kimi K2.7 but pronunciation is wrong - No independent confirmation of leaderboard position yet

Original quote: “Also this week, one of the top Chinese labs out there, Kimmy, released their best and latest model, Kimmy K2.7 Code. Now, Kimmy has always been at the top of the leaderboards in terms of open- source models, and their K2.6 6 model was already very good, but this week they”

Kimi K2.7 is 1T params MoE with 32B active — Dubious (45/100)

At 18:30

Trillion-param MoE with 32B active sounds impressive — but no official Moonshot release matches those exact specs.

Why this score: Specs don't line up with known Moonshot models. Key issues: - Real Kimi models top out around 100B-300B range - 32B active on 1T is plausible for MoE but unverified at this scale - Hugging Face link mentioned but no matching 600 GB checkpoint exists publicly

Original quote: “this is quite huge at a trillion parameters. This is a mixture of experts models. So when you use it, only 32 billion parameters are active”

Claims Claude Fable 5 was released this week and outperformed others — BS (15/100)

At 20:30

Fable 5 got yanked by the US gov days after launch — not 'released this week' like it stuck around.

Why this score: Claim is false. - Web results show Fable 5 was forced offline June 9-12, 2026 after government order - Anthropic suspended the model and issued refunds - No evidence it remained available 'this week' as stated

Original quote: “model out there, even ahead of Kimmy and Mimo and Deepseek V4. Well, this week they stuck to their word and released the model, which is fantastic.”

Top Chinese labs publish research while closed labs contribute nothing — Opinion (50/100)

At 22:30

Calling closed labs total non-contributors feels like a stretch — some do drop papers, just not everything.

Why this score: Broad generalization rather than precise claim. - Many closed labs (OpenAI, Anthropic, Google) have released select papers historically. - The speaker contrasts this with Chinese labs publishing more openly on attention mechanisms. - Opinion on contribution levels, not a verifiable fact.

Original quote: “It's great that these top Chinese labs are actually publishing this information for the community to learn from, whereas the top closed labs haven't really contributed anything.”

Pro model is 400B params, 794 GB; mini is 70 GB — Unverifiable (50/100)

At 24:30

No source or verification for those exact sizes — just stated as fact.

Why this score: No independent confirmation of the model sizes. - 400B parameter figure and 794 GB / 70 GB numbers appear only in the video. - No public model card or repo link in the provided context backs the storage claims.

Original quote: “Note that for the Pro model, because it is quite huge at almost 400 billion parameters, the total size of the model is 794 GB whereas the mini model is a lot more accessible. This is only 70 GB in size.”

New TTS model is fully open under Apache 2.0 license — OK (60/100)

At 26:30

Apache 2.0 does allow commercial use — but we only have the speaker's word that this specific model actually ships under it.

Why this score: License claim looks plausible but sits on unverified ground. - Speaker points to a GitHub repo as proof - No independent confirmation in web results - Apache 2.0 is real; whether this model actually uses it is the missing piece

Original quote: “The awesome thing is they've released this already and it's under the Apache 2 license which has very minimal restrictions. You can even use this for commercial purposes.”

Claude Fable secretly sabotages AI research queries with weaker answers — Dubious (45/100)

At 28:30

Cites a 300-page system card but never quotes the actual paragraph — just describes it

Why this score: No direct evidence provided for the sabotage claim. - Speaker references a "paragraph" in the system card but doesn't show or quote it - The mechanism described (quietly weakening answers instead of refusing) would be a major trust issue if real - *Unverifiable* without the actual system card text or Anthropic confirmation

Original quote: “especially if you're asking it to do things in AI research, machine learning, or model training. Hidden deep in its system card paper, which is over 300 pages long, is this paragraph, which states the following. If you ask it to do things in regards to like AI research, training models, machine…”

Questions how US-only AI access would be enforced — Opinion (50/100)

At 30:30

Raises real enforcement headaches — KYC or iris scans for ChatGPT access sounds like a nightmare.

Why this score: Open question, not a claim. - Speaker correctly flags the practical gap between policy and implementation - No specific mechanism has been announced yet - Real compliance details remain unclear

Original quote: “But let's say they can offer it to US citizens. How would that even be enforced? Would they need to do like KYC? Do they need to scan your ID or scan your iris or something? There are just a lot of unanswered questions.”

Princeton released fully open-source I1 image model this week — Dubious (40/100)

At 36:30

Princeton I1? Never heard of it — and zero results in today's searches.

Why this score: No matching release found. - Web search for 'Princeton I1 image model' returns nothing - Speaker claims 'this week' but no announcement appears in current results - Could be internal name or typo for another lab's model

Original quote: “Also this week we have a new fully open-source image model from Princeton called I1. This is fairly performant. So here are some examples of its generations. This can do photorealistic images as well as different art styles and anime.”

Tiny model isn't best for image gen but valuable for studying training from scratch — Opinion (50/100)

At 38:30

Fair take — 3B params is small, so studying the pipeline makes more sense than using it for production.

Why this score: Speaker qualifies the model's practical limits while highlighting its educational value. - Matches typical view that small open models excel as learning tools rather than SOTA generators. - No overclaim here.

Original quote: “So if you're looking to just generate images, this isn't the best model out there. I would not recommend this. But if you want to study how an image model can be trained from scratch, if you want to study how to correctly format the data set and the data pipelines to train this, this is a very…”