OpenAI Just Revealed Something More Dangerous Than AGI
Credibility score: 62/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
BSmeter analyzed "OpenAI Just Revealed Something More Dangerous Than AGI" and rated it 62/100 for credibility (a BS score of 38/100 — mostly credible), on 2026-09-17. Its weakest claim — "Dismissing quantity without acknowledging potential for emergent quality — a false dilemma." — scored 20/100 and was flagged as false dilemma. 27 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 27 claims analyzed: 2 scored under 40, 8 between 40 and 69, and 17 at 70 or above.
Claims analyzed
OpenAI is chasing recursive self-improvement, which is more consequential than AGI. — Confidence Mismatch (45/100)
At 0:02
He says OpenAI is 'openly' chasing this, but the 'more dangerous' part is his own spin. They're developing, not declaring a threat. 😈
Why this score: The speaker confidently asserts that recursive self-improvement is 'way more consequential' and 'more dangerous' than AGI, attributing this sentiment to OpenAI. While OpenAI is indeed researching recursive self-improvement, the framing that they 'openly said it's what they're chasing hardest' and that it's 'more dangerous' is the speaker's interpretation, not a direct quote or stated objective from OpenAI itself. It's a leap from research focus to existential threat, presented as OpenAI's own view.
Original quote: “Everyone's been watching for AGI, the moment AI gets as smart as us. But there's another line that could turn out to be way more consequential. And Open AI just openly said it's what they're chasing hardest. It's AI that helps build the next AI. That one helps build the one after it a bit faster…”
Claims recursive self-improvement is OpenAI's top focus and three teams have published systems for it. — No Frame (75/100)
At 0:35
He says it's the top focus and points to recent research. It's a straightforward statement of current AI development. 🔥
Why this score: The speaker is presenting information about OpenAI's priorities and recent advancements in AI research, specifically recursive self-improvement. This aligns with public discussions and reports regarding frontier AI capabilities and risks. It's a direct statement of fact, not a rhetorical trick.
Original quote: “This is number one, by a pretty wide margin. And this isn't just talk anymore. Over the past few days, three different research teams published systems that actually pull off pieces of it. One of them goes all the way down to retraining the model itself.”
Admits no proof that recursive self-improvement is more dangerous than AGI. — No Frame (75/100)
At 1:18
A rare moment of honesty, admitting the danger isn't 'proven.' Acknowledges the lack of definitive evidence. 😈
Why this score: The speaker is being transparent about the current state of evidence regarding the comparative danger of recursive self-improvement versus AGI. This is a direct and honest admission, not a manipulative framing technique. It sets up the subsequent argument as a potential risk rather than a certainty.
Original quote: “Nobody has proven that this is automatically more dangerous than AGI.”
Cites Reuters as treating recursive self-improvement as central to AI risk debate. — No Frame (75/100)
At 1:43
Citing Reuters for its coverage of AI risk. A direct, verifiable reference to a news source. 😈
Why this score: The speaker is referencing a reputable news organization, Reuters, to support the claim that recursive self-improvement is a significant topic in the discussion of AI risk. This is a verifiable statement about media coverage, not a rhetorical manipulation.
Original quote: “That feedback loop is why Reuters now treats recursive self-improvement as a central part of the debate over frontier AI risk.”
Claims Brown says AI models are 100 times better than humans at some bug-finding tasks. — Anonymous Authority (45/100)
At 2:13
A specific number, '100 times better,' attributed to 'Brown' without context on how that metric was derived. Sounds impressive, but it's just a quote. 💀
Why this score: The speaker attributes a very specific and dramatic statistic ('100 times better') to 'Brown' (Nome Brown, mentioned earlier). While Brown is identified as being 'behind OpenAI's reasoning models,' the context or methodology for this '100 times better' claim is completely absent. It's presented as a definitive fact from an authority without any supporting data or explanation of the measurement, making it an appeal to anonymous authority.
Original quote: “Brown [music] says current models are extremely diligent at it, and on some of that work, they're around a 100 times better than humans doing it manually.”
Claiming human judgment is the 'last big advantage' — a bold, unproven assertion. — Confidence Mismatch (45/100)
At 2:49
He says 'might be' then declares it 'the last big advantage.' That's a leap, not a conclusion. 💀
Why this score: The speaker presents a speculative 'might be' as if it's a definitive, almost singular, remaining human advantage. It's a dramatic statement without evidence to back its finality.
Original quote: “That kind of judgment might be one of the last big advantages humans still hold inside the AI research loop.”
Dismissing quantity without acknowledging potential for emergent quality — a false dilemma. — False Dilemma (20/100)
At 3:04
He sets up 'fast researchers' vs. 'knowing what's worth researching' as if they're mutually exclusive. Why not both? 😈
Why this score: The speaker creates a false dichotomy, implying that a large number of researchers (or AI agents) cannot also possess the 'taste' or judgment to know what's worth pursuing. It ignores the possibility of systems that combine both speed and strategic insight.
Original quote: “But numbers only get you so far. A thousand fast researchers aren't worth much if none of them knows what's worth researching.”
Introducing 'Dream RSI' and its creators — straightforward information. — No Frame (75/100)
At 3:19
Alright, they're actually naming the project and the institutions. A rare moment of clarity. 🔥
Why this score: The speaker is simply stating the name of a research project and listing the organizations involved. There's no apparent manipulation or rhetorical trickery here; it's direct information.
Original quote: “The first is called Dream RSI from researchers at Google, Google DeepMind the University of Maryland, and the University of Virginia.”
Describing current AI limitations where human-defined strategies are 'locked in place' — a factual observation. — No Frame (75/100)
At 3:47
He's just laying out how these systems currently work. No spin, just the setup. 😈
Why this score: This statement describes a common characteristic of current AI systems, where the overarching search strategy is often predefined by human engineers. It's presented as a factual limitation that the new approach aims to overcome, not as a manipulative claim.
Original quote: “But the strategy steering all of that, how many ideas to try at once, when to give up on a path, when to keep digging is usually written by humans and then locked in place.”
Explaining the 'Dream RSI' process — no tricks here, just setting the stage. — No Frame (75/100)
At 4:34
Just laying out how this 'dreaming' system supposedly works. Standard explanation. 😈
Why this score: The speaker is describing the core mechanism of the 'Dream RSI' system, explaining how it tests strategies without new experiments. This is foundational information for the rest of the segment, presented without any obvious rhetorical manipulation.
Original quote: “That's the trick. Dream RSI tests thousands of alternative strategies against the old record. One might drop a weak branch much earlier. Another might explore three paths at once. Another might push harder right when progress starts slowing down. Since every outcome is already recorded, none of…”
Claiming a 'built-in safety net' ensures no worse performance — a comforting thought with no real teeth. — Confidence Mismatch (45/100)
At 5:08
A 'safety net' that only guarantees 'not worse' isn't exactly a win, is it? That's a low bar for 'safety.' 💀
Why this score: The speaker presents the 'safety net' as a significant feature, but the guarantee is only that the new strategy won't perform 'worse' in replays. This is a minimal assurance, not a robust safety net, and the framing overstates its protective value. It's a confidence play, not a true safeguard against unforeseen issues in real-world application.
Original quote: “There's even a built-in safety net. The strategy already in use is always one of the candidates. So whatever gets sent back out can never score worse in those replays than what it replaced.”
Highlighting 'striking results' with specific, impressive metrics — but only two examples. — Cherry-Picked (20/100)
At 5:25
They tested it on 'eight tasks' but only gave you the two best numbers. That's not 'striking results,' that's a highlight reel. 🍒
Why this score: The speaker claims 'striking results' and mentions testing on 'eight tasks,' but only provides specific, positive metrics for two of those tasks. This is a classic cherry-picking move, presenting only the most favorable data points to support the claim of 'striking' performance, while omitting results from the other six tasks which may not have been as impressive.
Original quote: “The results are pretty striking. The team tested it on eight tasks from algorithm design to math optimization to writing the low-level code that helps graphics chips run AI faster. On one chip task, it matched the old approach with about two and a half times fewer attempts. On another, it scored…”
Presenting a 'most dramatic result' with a massive difference in calls — a stark comparison. — No Frame (75/100)
At 5:46
A 300 vs 51,000 call difference is genuinely significant. No trickery here, just a clear win. 🔥
Why this score: The comparison between Dream RSI's 300 calls and Simple TE's 51,000+ calls on a specific statistics problem is a direct and quantifiable metric. This stark difference, if accurate, genuinely represents a dramatic improvement in efficiency, and the speaker presents it without obvious rhetorical manipulation.
Original quote: “The most dramatic result came on a classic statistics problem where Dream RSI came out ahead using about 300 calls to its AI agent while a competing system called Simple TE's used [music] more than 51,000.”
Describing the system's adaptive strategy of 'holding back' then 'widening search' — a narrative of intelligent behavior. — Loaded Language (45/100)
At 6:00
It 'got more careful' and 'opened back up' — that's not just data, that's a story of intent. It's an algorithm, not a strategist. 😈
Why this score: The speaker uses anthropomorphic language like 'got more careful' and 'opened back up' to describe the algorithm's behavior. While the underlying data (attempts per round) might be accurate, framing it with terms implying conscious decision-making ('Hold back when things are working. Widen the search when you're stuck') imbues the system with a level of agency and intelligence that goes beyond a purely technical description of its adaptive parameters. This loads the description with emotional appeal, making the system seem more 'intelligent' than the raw data might suggest.
Original quote: “But the part that really got my attention is how the strategy behaved over time. On one chip task, the best score climbed from 0.427 to 1.898 across nine rounds. Early on, the system got more careful, cutting attempts per round from about 110 to 50. Then, when improvements started to stall, it…”
Describing a search strategy as a 'crude, narrow echo' of instinct. — No Frame (75/100)
At 6:30
Just setting the stage with a descriptive analogy. No trickery here, just a bit of poetic license. 😈
Why this score: The speaker is using evocative language to describe a concept, not making a factual claim that needs verification. It's a metaphorical comparison to human instinct.
Original quote: “things are working. Widen the search
when you're stuck. It's a crude, narrow
echo of the instinct Brown was
describing.”
Claiming that giving AI specific search advice made results worse, boxing it in. — No Frame (75/100)
At 6:37
A straightforward observation about AI behavior. Sometimes less guidance is more. 🔥
Why this score: This is a direct statement about an observed outcome in AI research. It's a finding, not a rhetorical device. The speaker is relaying information about how the AI performed under different conditions.
Original quote: “researchers tried the more obvious
route, summarizing past runs into
written advice about where to look,
results got worse, telling the AI where
to search just boxed it in.”
Defining the 'harness' as the AI's operational machinery beyond the core model. — No Frame (75/100)
At 6:59
Clearly defining a technical term. This is just information, not a trick. 😈
Why this score: The speaker is providing a clear definition of 'harness' in the context of AI agents, breaking down its components and functions. This is purely informative.
Original quote: “Around
it sits the rest of the car. The
machinery that runs the cycle of
thinking and acting, handles tools,
[music] processes what the agent sees,
manages its memory, and decides when a
job is actually done. Engineers call
that the harness.”
Explaining how Modular RSI isolates and fixes faulty parts of the AI harness. — No Frame (75/100)
At 7:14
Describing a technical process. It's a clear explanation of how the system works. 🔥
Why this score: This segment describes the methodology of Modular RSI, detailing how it identifies and addresses weaknesses in the AI's 'harness.' It's a factual description of a research approach.
Original quote: “Modular RSI splits the
harness into those five parts. It
compares runs where the agent succeeded
with runs where it failed on the same
task, hunts for weaknesses that keep
showing up across many tasks, and then
changes only the part that seems to be
at fault.”
Stating that harness parts evolve separately and then merge. — No Frame (75/100)
At 7:30
Just a continuation of the technical description. No hidden agenda here. 😈
Why this score: This is a straightforward description of the modular evolution process within the AI's harness. It's a technical detail, not a claim requiring external verification.
Original quote: “Each part evolves separately
and the upgrades get merged back into
one harness at the end.”
Presenting specific accuracy improvements on Terminal Bench 2.0 and SWEBench Verified. — No Frame (75/100)
At 7:57
Giving specific numbers for performance improvements. The data is right there. 🔥
Why this score: The speaker is citing specific, quantifiable results from the research, including percentage increases on two different benchmarks. These are direct data points from the study being discussed.
Original quote: “On
terminal bench 2.0, zero, a set of long
multi-step command line tasks. Accuracy
rose from 47.57
to 52.43%. On SWEBench verified, 500 real software
problems checked by humans. It went from
73.40
to 76.45.”
Claiming broad applicability for AI improvements based on specific benchmarks. — Confidence Mismatch (45/100)
At 8:34
They're showing specific percentage bumps and then leaping to 'it worked in different cars' and 'carried over' like it's a universal law. That's not how this works. 💀
Why this score: The speaker presents specific, modest percentage improvements on benchmarks (GLM 5.2, miniax 2.5) as evidence of broad, generalizable improvement across 'different cars' and 'tasks and models the system never evolved on.' The leap from specific benchmark gains to universal applicability is a classic confidence mismatch, implying a wider impact than the data directly supports.
Original quote: “Consistency improved too, which counts for a lot if you actually want to rely on an agent. It even worked in different cars. Dropped onto other AI models, the evolved harness took GLM 5.2 [music] from 59.55 to 61.80% 80% on terminal bench and miniax 2.5 [music] from 41.57 to 44.94. So these weren't…”
Introducing 'Science Buddy' as a deeper step in recursive self-improvement. — No Frame (75/100)
At 9:01
Just introducing a new project and its claimed methodology. Nothing tricky here, for once. 😈
Why this score: The speaker is introducing a new project, Science Buddy by PHAI Labs, and describing its core concept of 'recursive self-improvement.' This is presented as a straightforward description of the project's design philosophy, not an attempt to mislead or frame information deceptively. It's a clear statement of what the project aims to do.
Original quote: “Then PHAI Labs released Science Buddy and it pushes the idea one layer deeper. They call it recursive in recursive self-improvement. [music]”
Describing Science Buddy's capabilities and its recursive learning process. — No Frame (75/100)
At 9:13
Detailing the features and the feedback loop of Science Buddy. It's a description of how the system is designed to work. 😈
Why this score: The speaker is providing a detailed explanation of Science Buddy's functionality, including its tools, data handling, and the sophisticated feedback mechanism where it learns from scientist interactions. This is a descriptive account of the system's design and intended operation, without any apparent manipulative framing or misleading claims. It's explaining a technical process.
Original quote: “Science buddy isn't a benchmark experiment. It's a working research space for real scientists with 224 tools covering areas like genomics, cancer biology pharmarmacology bioiming and scientific literature. It runs code, handles data and biological sequences, and keeps a full record of what it did…”
Explaining the inner and outer loops of Science Buddy's self-improvement mechanism. — No Frame (75/100)
At 9:56
Breaking down the two-loop system for self-improvement. It's a technical explanation, not a trick. 😈
Why this score: The speaker is explaining the technical architecture of Science Buddy's recursive self-improvement, detailing the 'inner loop' for refining the harness and the 'outer loop' for generating new training material. This is a clear, step-by-step description of a proposed system design, presented as factual information about the project's methodology. There's no evident framing intended to mislead or manipulate.
Original quote: “From there, two loops kick in, one inside the other. In the inner [music] loop, the main model stays frozen. A separate fixed AI model studies failures and proposes small changes to the harness like better instructions, reusable skills, or different memory settings. Every change has to beat the…”
PHAI's evidence is only at the case study stage — a crucial caveat. — No Frame (75/100)
At 11:00
They're admitting the limitations of the evidence right up front. Rare honesty, I'll give them that. 😈
Why this score: The speaker explicitly states that the evidence for PHAI's advancements is still in the early 'case study stage' and across only 'four kinds of scientific work.' This is a direct acknowledgment of the current scope and limitations, which is a straightforward presentation of facts.
Original quote: “Now, PHAI's
[11:03] evidence is still at the case study
[11:05] stage across four kinds of scientific
[11:08] work, from reading literature to
[11:09] assessing gene variance.”
Not proof of intelligence explosion, but 'structurally' the same loop — a classic bait and switch. — Volume Game (45/100)
At 11:09
Dismisses 'intelligence explosion' then immediately pivots to 'structurally, this is the kind of loop.' They give with one hand and take with the other. 😈
Why this score: The speaker first states that the current evidence is 'not proof of an intelligence explosion,' which is a reasonable caveat. However, they immediately follow this by saying, 'But structurally, this is the kind of loop researchers have been talking about for years.' This rhetorical move downplays the current lack of evidence while still trying to connect it to the more dramatic concept of an 'intelligence explosion' through a 'structural' similarity, implying a future inevitability without present proof.
Original quote: “It's not proof
[11:12] of an intelligence explosion. But
[11:14] structurally, this is the kind of loop
[11:16] researchers have been talking about for
[11:18] years, and they're thinking much bigger.”
Admits no AI redesigns itself without humans, then immediately pivots to 'Brown's comment lands differently.' — Volume Game (45/100)
At 11:56
They tell you 'nobody has shown' autonomous AI, then immediately use that to say 'Brown's comment lands differently.' It's a setup for manufactured urgency. 💀
Why this score: The speaker clearly states that 'Nobody has shown an AI that keeps redesigning itself into something smarter with humans fully out of the loop,' and that 'Every system here runs inside fences with fixed judges, fixed budgets, and people still steering.' This is a direct admission of the current limitations. However, they immediately pivot to 'But that's also why Brown's comment lands differently right now,' using the acknowledged limitations as a springboard to imply greater significance or danger, despite the lack of fully autonomous AI. It's a classic volume game: state a truth, then…
Original quote: “Nobody has shown an AI that
[11:59] keeps redesigning itself into something
[12:01] smarter with humans fully out of the
[12:03] loop. Every system here runs inside
[12:06] fences with [music] fixed judges, fixed
[12:08] budgets, and people still steering.”
See the full analysis with timestamps →