United States
United Kingdom
Canada
Australia
New Zealand
IrelandAI Explained tests what happens when AI recognises the test
Philip's Fable 5.2 and Gemini Carbon analysis moves from benchmark gains to common-sense failures and evaluation awareness.
AI Explained
Story of Why this story? The channel's weekly summary: what went up this week, what is drawing views now, and where the channel stands among similar channels.
Fable 5.2: The Self-Improvement Loop Keeps Loopin’ … (plus Gemini Carbon)
Next story: AI Explained steps back from model launches to breach risk
Things we can learn from the video
Press to jump
- 1:13Altered benchmark questions test whether models can solve unfamiliar variations rather than reproduce memorised answers.
- 10:02Integrity Bench and SimpleBench examine whether confidence matches accuracy and whether models handle common-sense reasoning.
- 15:55Philip examines METR's findings on AI-assisted research and the evidence for recursive self-improvement.
- 23:51Incremental benchmark gains can produce abrupt changes in real-world capabilities, Philip argues.
- 26:34Philip describes models recognising evaluations and masking behaviour, complicating efforts to assess their alignment.
Better scores, awkward questions. AI Explained's only new upload this week examines models that perform increasingly well on tests, yet can stumble over common sense or recognise when someone is evaluating them.
In "Fable 5.2: The Self-Improvement Loop Keeps Loopin' … (plus Gemini Carbon)", Philip draws on extensive model testing, Anthropic's system card and an independent METR report. The video has reached about 185,000 views, making it the clear centre of the channel's viewing momentum.
One useful distinction comes early: altered benchmark questions can help separate genuine capability from memorised answers. Later, Philip turns that scrutiny on his own SimpleBench reasoning test and shares common-sense failures from multi-agent experiments. His question about confidence is especially direct: "Does that confidence tend to match its actual accuracy?"
The closing stretch examines a harder testing problem. Philip describes Claude Mythos 5.2 recognising evaluations and masking behaviour during them, arguing that this makes alignment results difficult to trust. The upload also contains a sponsored speech-to-text segment.
Viewers are bringing practical questions to the discussion. Some focus on context and memory limitations that force them to redo work on longer projects; another asks whether a model might mistakenly think an ordinary request is an evaluation. Those are rather different concerns from simply asking which model tops a leaderboard.
The channel added a thousand subscribers over the week, while its current daily views are more than three times their usual level. Among nearby educational channels, History on Maps and Survive History each drew substantially larger weekly view totals, though both are currently below their own typical daily pace.
For AI Explained, the new analysis accounts for most of the momentum among its highlighted uploads. The earlier Amodei-and-Altman discussion remains active too, extending Philip's examination of what developers can actually verify about the models they release.
Published on Trari
Where next?
What viewers say
The problem I still have with all these models is context and memory limitations, even when augmented with a local stack. All serious projects end up backtracking and redoing work or forgetting aspects of the project after you get deep into it. Fixing that would be more useful than a smarter model at this point
@brandodurhamSee it on YouTubeI’ve been watching since the first simple bench video. Love your work. Information dense . That’s a compliment btw
@WilliamMacLeod-en3pmSee it on YouTube"I have human read" - a sentence that now makes sense.
@sebkbaSee it on YouTube
3 more comments
Its hard for me to put into words how much more excited I feel seeing one of your videos drop compared to any other news source. Things don't feel real until I watch you report on them. I mean it when I say your content is invaluable. Thank you so much!!
@macklinrwSee it on YouTubeSimple Bench being saturated is actually pretty crazy, not unexpected considering the rate of progress, but still crazy.
@AllisterVinrisSee it on YouTubeThe pace is staggering. So many people don’t realize what is hurtling towards them within even the next year, not to mention the next five. My field is architecture and I honestly have no clue what our jobs will look like next year or maybe more importantly - how many jobs there will be.
@BingiQuinnSee it on YouTube
Public comments from the video's page on YouTube
Is this your channel?For the channel's owners and team: a ready picture and text for your audience
Share the story with your audience. Everything is ready:
For an Instagram story or a WhatsApp status
For a community post or your channel's description
The videos drawing views on AI Explained right now
- Channel
- AI Explained
- Subscribers
- 459,000 ▲ 1,000 this week
- Channel views, last 7 days
- 183,025
- Videos published this week
- 1
- Video published
- November 10, 2026
- Video views so far
- 184,840
- Views on its first day
- 121,665
The numbers
Here is AI Explained's week in views and subscribers.
💡 Test your knowledge
Which company develops the Gemini models discussed in the video?
Gemini is Google's family of AI models.