Tokens can make you rich, just do this – Mario Zechner
Credibility score: 50/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "Tokens can make you rich, just do this – Mario Zechner" and rated it 50/100 for credibility (a BS score of 50/100 — mixed credibility), on 2026-05-20. Its weakest claim — "New AI model gives 97% performance at 1/10th the cost" — scored 35/100 and was flagged as dubious. 30 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 30 claims analyzed: 3 scored under 40, 25 between 40 and 69, and 2 at 70 or above.
Claims analyzed
Tokens favor the rich; most people just chasing trends, slop code + agents let 6 people beat 50-100 person teams — Opinion (50/100)
At 0:00
Classic 'AI lets tiny teams crush big ones' take — heard it since GPT-3 days.
Why this score: Opinion piece on AI tooling.
- No data given on actual team size wins.
- 'Slop code' point is real but ignores long-term maintenance costs.
- Trend-chasing observation is fair but anecdotal.
Original quote: “think it's going to be a richman's game. People who have the means of production in the sense that they can afford the tokens have a massive edge. >> I've noticed there's a lot of people who don't actually believe in the fundamental technology. They're just here like for the trend. >> They try a…”
Been using LLMs since 2022 right after GPT launch — Personal Story (70/100)
At 0:30
Timeline checks out — GPT-3.5 hit public in late 2022
Why this score: Personal timeline matches known history.
- GPT-3.5 public release was November 2022
- Speaker's stated start date aligns exactly
Original quote: “I've been using LLM since basically 2022 when OpenAI came out with Jet GPD. Uh because I had a background in LLP and machine learning in the 2000s. Kept up with the field and was quite enticed by LLMs because they could basically do all the natural language processing tasks that I was painstakingly…”
Claude Code is not good but many people use it anyway or are used to slop — Opinion (40/100)
At 2:30
Opinion dressed as insider take — no data on how many are "used to slop".
Why this score: Pure opinion on Claude Code quality. Key points:
- Calls it "slop" and says it's not good
- Admits many people still use it
- No metrics or user data offered to back the "used to slop" idea
Original quote: “I think they're dog fooding uh what they're preaching and I think that's perfectly fine. They're kind of like a canery in a coal mine where you can see how good agents autonomously can work at the moment and if the products they spit out are any good. I personally don't think cloud code is good.…”
Anthropic cleared Claude's thinking traces after 1 hour idle to cut latency — Verified (85/100)
At 4:30
✅
Why this score: - Direct quote from Anthropic changelog dated March 26th
- Matches known Claude API behavior around context management
Original quote: “what does it say on March 26th we shipped the change to clear cloud Claude's older thinking from sessions that had been idle for over an hour to reduce latency. So if you had an agent run for an hour or you stepped away because you grabbed lunch and then came back, all the old thinking traces were…”
Removing thinking traces after 1 hour is like lobotomizing the model — Opinion (50/100)
At 4:52
Metaphor lands but overstates impact
Why this score: - Strong language for a routine context cleanup
- Real issue exists but "lobotomizing" exaggerates standard engineering practice
Original quote: “And that's like super bad because that's kind of like lobotomizing the model for that session if you take away half the information that it generated throughout the the session.”
Full control over context is the main goal for reliable coding agents — Opinion (50/100)
At 5:15
Fair take on agent reliability
Why this score: - Core point about consistency across sessions
- Reflects real pain point for long-running coding agents
Original quote: “So it comes down to context ultimately, right? Everything's about context really. Yeah. Yeah. So like that was one of the main things that you were seeking is like full control over the system problem. Making sure that you know your assistant is the same today as tomorrow.”
AI model hype fades due to psychological honeymoon period — Opinion (50/100)
At 6:31
Classic post-honeymoon disappointment — most of us have lived it 😂
Why this score: Personal experience framed as general pattern. Reasonable take but unquantified.
Original quote: “I have never felt that. So, I think there's two two things that are happening. The first one is psychological. A new model comes out. Uh you get super excited. You bang out a trillion tokens and you build new green field projects and like wow, it's amazing.”
Most perceived AI degradation is just lost enthusiasm — Opinion (50/100)
At 6:43
Explains the feeling but skips real model drift reports 🧐
Why this score: Frames drop-off as mindset shift. Plausible contributor but not proven as primary cause.
Original quote: “Um and then eventually you put it to work for something that's more important and you see ah [ __ ] this still doesn't work and I hear it [ __ ] up and so on and so forth. So I think it's mostly psychological in in that sense that the initial uh enthusiasm vanishes as you start using the model for…”
Most AI issues come from changing harnesses, not model changes — Opinion (50/100)
At 6:55
Blames the wrapper more than the model — fair but unproven 🤷
Why this score: Highlights infrastructure churn. Hard to isolate without controlled tests.
Original quote: “Um and the second part is well I mean it's also the harness specifically cloud code. It changes so much that it breaks things just by changing the harness not the model. I personally would attribute most of that to those two things.”
AI model degradation is mostly psychology, not real — Opinion (50/100)
At 8:30
Calls collective doubt a mass delusion — classic "you're imagining it" move.
Why this score: Opinion, not a testable fact. Key points:
- Speaker frames reported degradation as "psychology" and "collective psychosis"
- No data or studies cited to back the dismissal
- Personal take on why others report issues
Original quote: “I think if you talk to open code people that use entropic models, they will probably tell you the same that they don't feel this degradation as much or at all. Um so yeah, I I think this is mostly psychology.”
Past degradation was just a one-off bug that got fixed fast — OK (55/100)
At 8:46
Acknowledges real bug but downplays it as quick fix — plausible but light on details.
Why this score: Mildly credible but thin. Key issues:
- Admits actual degradation occurred from inference bug
- No specific date, model, or source named
- General claim that teams "feel immediately and fix" lacks evidence here
Original quote: “To be fair, they had a bug at some point earlier this year where they [__] up some code in a kernel or something where there was actual model degradation because of a bug in the inference engine, but that's stuff that they usually themselves u feel immediately and fix.”
AI lets him do projects he couldn't before and test fast — Personal Story (50/100)
At 9:15
Classic dev productivity flex — sounds real but impossible to verify.
Why this score: Personal anecdote. Key points:
- Speaker describes his own productivity gains
- Valid individual experience, not a general claim
- No broader data offered
Original quote: “I love that I can be much more productive and can do projects that I previously wouldn't have done because I just wouldn't have the time to do them. I also love that it allows me to experiment with things real quick.”
Hype around Ralph loops is tiring and mostly misunderstood — Opinion (50/100)
At 9:50
Calls out hype fatigue — hard to argue with that part.
Why this score: Opinion on AI marketing. Key points:
- Speaker tired of overhyping
- Distinguishes real scientific loops from PRD/spec iteration
- "I've never seen this work" is his experience, not data
Original quote: “The thing I hate is the [__] hype. It's just so tiring. Things like Ralph loops. Ralph loops can work, but people misunderstand why and how they work.”
Dark factories and similar AI hype don't actually work — Opinion (50/100)
At 10:30
Straight call-out of performative AI theater.
Why this score: Speaker is giving his direct opinion on current AI practices rather than stating verifiable facts. This is a value judgment about whether certain approaches qualify as 'cargo culting.'
Original quote: “I've never seen this work. If people make it work, more power to them. I For me, it's cargo culting.”
Insiders at AI Engineering Europe admitted hype doesn't work but sell it anyway — Personal Story (50/100)
At 10:50
Personal conference story — unverified but plausible.
Why this score: This is the speaker recounting private conversations he claims to have had at a specific event. No public record or corroborating sources exist to confirm or deny what anonymous attendees allegedly said.
Original quote: “Two weeks ago, I was in London at AI Engineering Europe, a big conference, and behind the scenes, a lot of people would come talk to me and be like, 'Yeah, all of this [__] doesn't work, like dark factories and so on, but we still need to sell it.'”
AI tokens will create a rich vs poor divide — Opinion (50/100)
At 12:30
Classic inequality worry — but every tech wave starts expensive.
Why this score: Opinion on economic access. No hard data presented, just forward-looking view.
- Token costs are dropping fast in practice.
- Past tools (computers, internet) followed similar early-price patterns.
Original quote: “I definitely think it's going to be a rich men's game. the people who have the means of production in the sense that they can afford the tokens have a massive edge over people who cannot afford the tokens.”
Wife's output 5xed after two nights with AI coding agent — Personal Story (55/100)
At 14:30
Classic one-person success story — no controls, no replication.
Why this score: Single anecdote with no supporting data.
- Claims 5x scientific output from one non-technical person
- No metrics, time periods, or comparison shown
- Common pattern in AI marketing where one impressive case is presented as proof of broad impact
Original quote: “>> I 100% agree with you. My my first observation of that was last summer when I gave my wife um she's a linguist. She's non-technical. I gave her uh a coding agent because she's working in on on linguistic data, lots of it. Like she does interviews with people, transcribes that annotates the…”
New AI model gives 97% performance at 1/10th the cost — Dubious (35/100)
At 16:30
97% same intelligence at 1/10th cost sounds magical — needs receipts.
Why this score: No specific model named as "deep v4". Performance claims at that level usually come with trade-offs in speed, context length, or edge cases. Real cost drops exist but rarely hit 90%+ parity at 10x cheaper without major caveats.
Original quote: “especially after yesterday's release of deep v4 we're starting to probably see a trend of optimizing for these different workflows where like you can have the you know 97% of the same capacity same intelligence for like onetenth 17th of the cost”
Anthropic's inference margins are 70% or higher — Dubious (35/100)
At 18:30
70% margin number dropped with zero source — sounds impressive but feels pulled from thin air 💀
Why this score: No public data backs the 70% figure. Key problems:
- Anthropic's actual financials are private
- Industry inference margins typically range 30-50% based on known cloud costs
- Speaker presents it as insider knowledge without evidence
Original quote: “I know the margins at least on inference for the US companies and anthropic 70%. Yeah. So yeah, it can even be more than that.”
Anthropic has stronger brand than OpenAI in enterprise — Opinion (50/100)
At 20:30
Subjective brand ranking — no data shown.
Why this score: Pure opinion on brand perception. No metrics, surveys, or market share data provided to back it up.
Original quote: “they have a perfect story. Um OpenAI, not so much. They're trying to get there, but I think Entropic just has a really good brand”
OpenRouter's $200 tier gives far more usage than Anthropic's equivalent — Opinion (45/100)
At 22:30
Depends entirely on his usage patterns — could be true for him, meaningless for others
Why this score: This is a comparative opinion about API pricing tiers. Value perception varies heavily by workload, so it's inherently subjective rather than factual.
Original quote: “if you compare these tiers at open to infropic you really feel the difference like with I I never hit my limits on $200 codeex and I always run multiple in parallel with close code it's like easy to do that in a day or two you can easily hit your limits”
US poaches AI talent from Europe and China due to money — Opinion (50/100)
At 24:30
Classic brain-drain narrative — sounds plausible but glosses over Europe's own choices.
Why this score: The speaker frames it as simple money-driven poaching.
- Europe has fewer big AI exits and weaker VC ecosystem compared to US.
- Many top European AI researchers do move to US labs, but local factors like university spinouts and funding play roles too.
Original quote: “Yeah, it's mostly America poaching the talent from both Europe and China actually. I mean it's not like there's not US native people who who are also amazing AI people obviously but the talent is fleeing to the US because that's where the money is.”
Used Open Claw AI to build custom calorie tracking files — Personal Story (50/100)
At 28:30
Classic personal workflow flex — sounds handy in theory.
Why this score: This is a personal anecdote about using an AI tool to create custom tracking files instead of downloading an app. No broader claim is being made about effectiveness or scalability, so it sits in the realm of individual experience rather than verifiable fact.
Original quote: “my first thought was using Open Claw to just create a CSV file and log my calories and then I created a second file which is Markdown file with approved meals and you know it has like the exact macro for each meal”
Only this AI integrates new research papers instantly — Sketchy (35/100)
At 30:30
Plenty of apps and researchers already pull latest papers — not unique.
Why this score: Overstates exclusivity. Tools like PubMed, Perplexity, and research assistants already ingest new studies. No proof this setup is faster or more accurate.
Original quote: “if I read a new paper on some dietary stuff or sports stuff I can immediately integrate that you know and no app in the world does that”
Believes Jevons paradox means AI agents boost productivity without replacing workers — Opinion (50/100)
At 32:30
Classic tech optimism — history says productivity gains rarely save all jobs 💼
Why this score: Opinion on AI future.
- Draws parallel to Jevons paradox but applies it to software agents.
- Ignores historical patterns where automation reduced certain roles even as output rose.
- Personal preference for non-dystopian outcome stated openly.
Original quote: “I'm big believer in Jeffron's uh paradox where agents don't replace people agents just make people even more productive.”
Seniors + AI agents will replace two juniors in any knowledge work — Dubious (40/100)
At 34:30
Big swing from 'some roles' to 'all knowledge work' with zero data 💀
Why this score: Sweeping generalization with no supporting numbers or studies. Key issues:
- Current AI tools still need human oversight in most professional fields
- No evidence yet for 2:1 replacement ratios across knowledge work
- Adoption rates vary wildly by industry and company size
Original quote: “a lot of juniors will not get any jobs because seniors plus an agent can now kind of replace and I'm not talking about software. I'm talking about all kinds of knowledge work. A senior in a knowledge work position can now replace two juniors with an agent um and still have their own output.”
Companies avoid juniors; pairing them with AI agents + seniors fixes it — Opinion (50/100)
At 36:30
Optimistic take on AI-assisted onboarding — real results still mixed.
Why this score: Junior + AI agent + mentor model is promoted by some firms but adoption is early-stage. Productivity gains exist in narrow tasks, yet full ramp-up time remains significant across most teams.
Original quote: “the problem is more that enterprises don't want to hire young people because usually takes a long time to to get them up to speed and to become productive. I think that's wrong. I think if as to at least in software engineering a junior plus an agent plus a senior who can mentor that junior perfect.”
Architecture and system thinking will matter more than syntax; learn it by hand first — Opinion (50/100)
At 40:30
Solid take — but the 'by hand' warning feels like gatekeeping nostalgia.
Why this score: Opinion piece on shifting developer skills.
- Speaker correctly notes that high-level design stays valuable when AI handles syntax.
- The 'you must do it by hand' part is classic advice that ignores how many senior devs already learned via tools.
- No data presented on whether current juniors are actually skipping fundamentals.
Original quote: “>> So it sounds like the skill set of software design and you know code with architecture is going to be more and more important than specific coding. >> Oh yeah, definitely. Like I don't care for syntax anymore, which is very nice. I can also use languages I haven't used in a long time or I've…”
AI agents will copy mostly bad old code from the internet — Opinion (50/100)
At 42:30
Classic LLM training gripe — no numbers, just vibes.
Why this score: AI training data is indeed full of flawed code. Key points:
- Public GitHub repos contain plenty of outdated or low-quality examples
- Models interpolate from existing patterns rather than inventing clean architectures
- Human oversight still beats blind trust in generated code
Original quote: “it learned all of that from the internet. And on the internet, it's my old shitty code, the old shitty code of a lot of other old people. And it's [__] Like 90% of that code is [__] And that is what your agent will propose the architecture and system should be. Don't let it do that. do it yourself,…”
90% of code in AI training data is garbage — Dubious (40/100)
At 44:30
Big round number with zero sources cited. Classic move.
Why this score: The speaker presents the 90/10 split as established fact, yet offers no dataset, paper, or measurement to support it. Training data quality varies by source and language, and no major public study has published this exact ratio.
Original quote: “And that's also the problem with code. Like 90% of the code that is in the training data is garbage. And the 10% which are pristine, awesome, amazing. They exist, but they're”
See the full analysis with timestamps →