It’s starting…
Credibility score: 46/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "It’s starting…" and rated it 46/100 for credibility (a BS score of 54/100 — mixed credibility), on 2026-06-05. Its weakest claim — "Anthropic has been secretly using unreleased 'Mythos' model since Q1" — scored 20/100 and was flagged as sketchy. 26 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 26 claims analyzed: 4 scored under 40, 22 between 40 and 69, and 0 at 70 or above.
Claims analyzed
Anthropic says AI self-building means slow development because society isn't ready — Opinion (50/100)
At 0:00
Calls the 'slow down' plea self-serving right after quoting Anthropic — no receipts yet, just the jab.
Why this score: Opinion framing.
- Speaker labels Anthropic's stance as self-interested without citing their actual paper or statements.
- The two points (unprepared society + slow down) are presented as Anthropic's direct conclusion.
- No external data offered to test whether the call to pause is genuine or competitive positioning.
Original quote: “AI is now literally building itself. And according to Anthropic, that means two things. One, society is not prepared. And two, we should actually slow down development. That is incredibly self-serving by anthropic.”
AI will soon build its own successor systems autonomously — Opinion (50/100)
At 0:30
Calls it a clear trend but admits recursive self-improvement isn't inevitable yet — hedging before the pitch lands.
Why this score: Future projection, not current fact.
- Speaker frames delegation as an existing trend
- "Given enough compute" is doing the heavy lifting
- No timeline or evidence provided for full autonomy
Original quote: “We are delegating a growing share of AI development to AI systems. Given enough compute, that trend points to an AI system capable of fully autonomously designing and developing its own successor.”
Predicts AI agents will soon build and train their own models, closing the loop on Claude improving itself — Dubious (35/100)
At 2:30
Recursive self-improvement is still sci-fi territory — current agents can't train frontier models.
Why this score: Speculative leap with no current evidence.
- No public system in 2026 can autonomously design, train, and iterate frontier-scale models.
- Training runs require massive infrastructure, data pipelines, and human orchestration.
- The speaker frames it as a near-future possibility without supporting milestones.
Original quote: “In the future, agents could become capable enough to build and train models themselves. If this happens, future versions of Claude could be continuously improved by Claude itself.”
Once recursive self-improvement arrives, humans are removed and compute becomes the sole bottleneck — Opinion (50/100)
At 4:30
Assumes full automation of AI R&D is imminent and removes every human role — that's still science fiction, not engineering.
Why this score: Speculative framing. Key issues:
- No current system demonstrates closed-loop recursive self-improvement without human oversight
- Even frontier labs still need researchers for architecture, evaluation, and safety decisions
- "Humans out of the loop" is a prediction, not a measured milestone
Original quote: “And as you can see here, the only bottleneck is compute. How much compute can you throw at the problem? How much parallelization can you have?”
AI agent task length doubling every 4 months, acceleration accelerating — Dubious (45/100)
At 6:40
Anthropic said tasks are getting longer — speaker turned that into 'acceleration accelerating.'
Why this score: Framing leap.
- The 4-month doubling claim comes from one internal Anthropic report
- 'Acceleration is accelerating' is the speaker’s interpretation, not their wording
- No public data confirms the prior 7-month baseline or continued trend
Original quote: “The length of tasks that they can reliably complete on their own that is AI agents has been doubling roughly every four months up from an earlier trend of doubling every seven months. That is the definition of accelerating. The acceleration is accelerating.”
Hints of recursive self-improvement already visible — Opinion (50/100)
At 7:00
Anthropic explicitly said it’s not here — speaker still calls the current signals 'hints.'
Why this score: Careful hedge ignored.
- The article states recursive self-improvement is not occurring
- Speaker reframes normal capability gains as early signals
- This is interpretive spin, not a factual claim
Original quote: “Now, they are very clear in this article that recursive self-improvement is not here yet, but we are seeing hints of it.”
Claude Opus 3 handled 4-minute human software tasks in March 2024 — OK (60/100)
At 7:27
The 4-minute figure is from Anthropic’s own internal eval — real but narrow.
Why this score: Limited scope.
- Anthropic measured this on a specific set of short coding tasks
- Does not mean general software engineering capability
- Time-to-complete for the model itself is not disclosed
Original quote: “March 2024, Claude Opus 3 could complete software tasks that take humans about 4 minutes to complete.”
AI task length jumped from 4 min → 90 min → 12 hours in 3 years — Dubious (45/100)
At 8:30
No public benchmarks show Opus 4.6 or Sonnet 3.7 hitting those exact time horizons — numbers appear made up.
Why this score: No public evidence supports the specific progression claimed.
- Anthropic has not released any model called Opus 4.6 or Sonnet 3.7 as of June 2026.
- Current frontier models are still measured in minutes for complex tasks, not 12-hour autonomous runs.
- The timeline appears to be an extrapolation rather than documented performance.
Original quote: “back then opus 3 could do it and four minutes is not that long but again just like a year previous all we had was tab complete so like uh a model being able to complete something that takes a human 4 minutes was actually a big deal but we've obviously come quite far since Then a year later, Sonnet…”
AI paper reproduction jumped from 20% in 2024 to nearly 100% now — Dubious (45/100)
At 10:30
No public benchmark or paper shows 100% success on novel AI research reproduction — that number appears invented.
Why this score: No supporting source for the 20% → 100% leap.
- Speaker gives no study name, dataset, or date for either figure
- Real reproduction benchmarks (like those from OpenAI or Anthropic) still hover well below saturation on truly novel ideas
- "Nearly 100%" is a massive jump with zero cited evidence
Original quote: “AI systems went from succeeding at reproducing the results roughly 20% of the time in 2024, 2 years ago, to saturating the benchmarks 15 months later. AI models can now successfully reproduce AI research paper novel ideas successfully nearly 100% of the time.”
In AI, execution is now easy and ideas are the hard part — Opinion (50/100)
At 10:47
Classic "ideas are the bottleneck" take — sounds insightful until you remember most labs still struggle with both.
Why this score: Opinion dressed as observation.
- Execution (training runs, infra) remains extremely expensive and technically hard
- Top labs still compete heavily on both ideas *and* engineering talent
- The claim flips the usual narrative without new evidence
Original quote: “This is the difference between having a new novel idea and knowing how to verify it, having taste in research, taste in a project that you're building versus actually going to execute that project, going to actually build it. We are at the point in AI right now in which the actual development is…”
Claude now solves highly underspecified problems on its own — Opinion (50/100)
At 12:30
Sounds impressive until you realize he's still the one picking the problems.
Why this score: Opinion on capability boundaries.
- Speaker acknowledges humans still define the task
- Claims models handle vague prompts effectively
- Distinguishes this from choosing what to build next
Original quote: “We've now gotten to the point where Claude can be handed an underspecified problem and figure out how to solve it. We are still handing the problem to Claude. The problem is still the realm of humans to determine. But even a very unspecified problem can be determined and developed by Claude.”
80%+ of Anthropic's merged code written by Claude as of May 2026 — Unverifiable (50/100)
At 14:30
Internal company metric with zero public data — just taking their word for it.
Why this score: No independent verification available for internal code authorship stats.
- Claim relies entirely on company self-reporting
- No public audit, GitHub data, or third-party confirmation exists
- Current date is June 2026 so the May 2026 timeframe is current but still private
Original quote: “As of May 2026, more than 80% of the code we merged into Anthropics codebase was authored by Claude. Now, this is super interesting.”
Anthropic has been secretly using unreleased 'Mythos' model since Q1 — Sketchy (20/100)
At 16:30
Claims a hidden frontier model called Mythos that no one else has seen.
Why this score: Zero corroboration.
- No public record or leak of an Anthropic model named Mythos.
- All other frontier model names are accounted for.
- Pure speculation dressed as insider knowledge.
Original quote: “They're using Mythos internally. And I think about why they might be doing that... They have been using Mythos since Q1. They're using this Frontier model and they did not release it publicly”
Anthropic banned XAI from using their models Jan 2026 — Unverifiable (50/100)
At 18:30
Story attributed to Kylie but no primary source or confirmation here.
Why this score: No independent verification available in current sources.
- Relies on single unlinked report from January 2026
- No public statements from Anthropic or XAI confirming cutoff
- Web context silent on this specific access revocation
Original quote: “Anthropic cut access to XAI... January 9th, 2026. XAI staff had been using Anthropics models internally through Cursor until Anthropic cut off the startup's access... they said, "No, no, you can't use our models to code your models"”
Anthropic poll: 4x output but 8x code means AI code is half as valuable — Dubious (45/100)
At 20:30
Poll of 130 people is real but the 'half as valuable' math is pure assumption — no quality data given.
Why this score: Poll numbers appear plausible but the leap to 'half as valuable' doesn't hold.
- Speaker cites internal March 2026 Anthropic survey — no public source linked.
- Equates quantity of code with value without any bug-rate, review-time, or deployment-success metrics.
- 'Eight times as much code' figure is stated without evidence in the clip.
Original quote: “In March 2026, in a poll, 130 employees from across anthropic research teams, the median respondent estimated that they produced around four times as much output with Mythos preview as they would have without access to any AI model. That is a significant improvement, four times your productivity.…”
Claims AI code preview is 50% less productive than humans — Unverifiable (50/100)
At 22:30
Drops the '50% less productive' number with zero source or study attached
Why this score: No supporting data provided. Key issues:
- No study, internal metric, or benchmark named
- 'Code mythos preview' appears to be an internal tool not publicly documented
- Can't verify without the actual measurement method or sample size
Original quote: “this code mythos preview is 50% less productive than human code”
AI writes 8x code so you need 8x reviewers — Sketchy (30/100)
At 24:30
Assumes linear 1:1 review scaling — ignores AI review tools and smarter processes.
Why this score: Linear scaling assumption is flawed.
- Real teams use AI reviewers and better tooling
- Doesn't account for code quality differences or automated testing
- The 'eight times' figure itself is presented without source
Original quote: “Because if AI is writing eight times as much code, you need technically eight times as many people to review that code.”
Claude-written code was worse than human code in late 2025, now at par, will beat humans soon — Opinion (50/100)
At 26:30
Staff "believe" AI code quality is catching up — that's a vibe check, not a benchmark.
Why this score: Anecdotal consensus among Anthropic employees. Key points:
- No public data or external audits backing the quality comparison
- Timeline prediction ("within the year") is speculative
- Speaker frames it as hopeful expectation rather than measured result
Original quote: “There isn't full consensus among staff at Enthropic, but many believe that the clawed written code was still worse in quality than human written code at Enthropic in late 2025 and is roughly at par. We expect it to be better within the year.”
Mythos preview hit 52x speedup vs Opus 4's 3x one year earlier — Unverifiable (50/100)
At 28:30
52x vs 3x sounds dramatic but no public benchmark or paper backs those exact numbers.
Why this score: No independent verification exists for these specific speedup figures.
- Speaker cites internal Anthropic tests with no published methodology
- Different models, different tasks, different baselines make direct comparison shaky
- Web search turns up zero corroborating sources for either the 3x or 52x claims
Original quote: “In May 2025, Opus 4 averaged 3x speed up over the starting code. In April 2026, brace yourselves, Mythos preview was achieving 52x”
Claude AI improved from 22% to 64% on experiment decisions since March 2024 — Unverifiable (50/100)
At 30:30
No public benchmarks or papers exist for "Claude Mythos preview" — that name doesn't match released models.
Why this score: No matching model name in Anthropic's releases.
- Claude 3 Haiku launched March 2024, but no 22% figure published
- "Mythos preview" has zero external references as of June 2026
- Internal-only metric can't be verified externally
Original quote: “Claude Haiku 3 back in March 2024, 22% of the time it would have done better. Fast forward to Claude Mythos preview, it is now at 64% of the time.”
AI now handles execution so ideas become the scarce valuable part — Opinion (50/100)
At 32:30
Classic pivot: used to be execution was everything, now AI flips it so the idea is king — neat framing but still just speculation.
Why this score: Opinion on shifting value in AI era.
- Speaker notes code generation is now fast enough that human review could bottleneck progress
- Argues this inverts the old Silicon Valley view that ideas are cheap and execution is the grind
- No data or examples given; it's a conceptual trade-off discussion rather than a testable prediction
Original quote: “On my team, there was this notion for a long time in Silicon Valley that ideas are cheap, right? Anybody could have the next billion dollar idea. The hard part is actually going out and building it. ... But execution, which was the hard part, can now be done by artificial intelligence, at least so…”
Even without AI taste, human prompting will create compounding acceleration — Opinion (50/100)
At 34:30
Assumes prompting alone multiplies output forever — no proof the bottleneck stays human.
Why this score: Opinion on future dynamics.
- Relies on assumption that direction-setting stays the hard part
- No data shown on how much actual acceleration prompting delivers today
- Treats current bottlenecks as permanent
Original quote: “So even if we suppose that Claude never achieves good taste, good research taste, a conservative reading of our evidence still implies compounding acceleration.”
100-person AI company equals 10k-100k person org — Dubious (45/100)
At 36:30
Said 100 people replace 100,000 — no data shows that ratio anywhere yet.
Why this score: No empirical basis for the 1,000x productivity leap.
- Current AI automation studies show 2-5x gains in narrow tasks, not 1,000x.
- The speaker offers no source or benchmark for the 100-to-100k figure.
- This is classic exponential hype without supporting measurements.
Original quote: “a 100 person company can do the work of a 10,000 or even a 100,000 person organization”
Energy owners at RSI moment freeze society in place forever — Sketchy (35/100)
At 38:30
Assumes whoever holds energy infrastructure at one instant locks out everyone else — real grids and markets don't work that way.
Why this score: Overstates lock-in effect. Issues:
- Energy markets allow new entrants, new sources, and policy shifts
- Compute is already moving to renewables and modular nuclear, not static monopolies
- Historical precedent for rapid infrastructure change contradicts the 'freeze' narrative
Original quote: “It might even be at the point at which that happens. It's not about capital. It's whoever had the compute, had the energy production mechanisms to begin with. Whoever had it at that moment, it's just all of society freezes.”
Anthropic publicly says AI development should slow down — Opinion (50/100)
At 40:30
They frame it as safety concern while sitting in first place — classic frontrunner logic.
Why this score: Opinion on policy preference.
- Speaker accurately quotes Anthropic's stated position
- No factual claim being made here, just interpreting their public stance
- The "slow down" line is directly from their safety report
Original quote: “If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing. They are saying, "We think we should slow down. We think we should slow down."”
Calls Anthropic's pause essay pure fear-based self-promotion — Opinion (50/100)
At 42:30
Calls their essay fear-based marketing — when the actual text was just one conditional sentence about pausing.
Why this score: Opinion piece. Speaker frames Anthropic's 2023 essay as cynical PR without evidence they wrote it only after securing frontier status.
- The essay predates their current model leadership
- No internal docs show the timing was strategic
- Classic "they're only moral because they're winning" argument
Original quote: “Fearbased marketing. I just don't get it. It it it hurts to read this stuff because it's like it's so self-serving from anthropic.”
See the full analysis with timestamps →