Deepseek just did the impossible
Credibility score: 62/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
BSmeter analyzed "Deepseek just did the impossible" and rated it 62/100 for credibility (a BS score of 38/100 — mostly credible), on 2026-09-18. Its weakest claim — "Setting up a straw man to knock down — a classic move." — scored 20/100 and was flagged as straw man. 52 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 52 claims analyzed: 1 scored under 40, 21 between 40 and 69, and 30 at 70 or above.
Claims analyzed
Deepseek's new model is the 'most optimized, efficient, and frictionless' AI model — a bold, sweeping claim. — Confidence Mismatch (45/100)
At 0:00
Calling it the 'most' of anything without a single metric or comparison is just pure hype. That's not data — that's a sales pitch. 😈
Why this score: The speaker uses superlative language ('most optimized, efficient, and frictionless') without providing any specific benchmarks, comparative data, or criteria to back up these claims. It's a confident assertion that lacks the necessary evidence to be truly verifiable, relying instead on strong adjectives to persuade.
Original quote: “Deepseek did it again. They've released a new model and not only is this among the frontier models out there, but it's also the most optimized, efficient, and frictionless AI model we've seen so far.”
The design of Deepseek's model is 'unexpected' and 'seems wrong' but is 'absolutely brilliant' — a classic setup. — Loaded Language (45/100)
At 0:13
He's building up the drama, calling it 'wrong' then 'brilliant.' That's not analysis — that's a storyteller setting the hook. 🔥
Why this score: The speaker uses emotionally charged language to create a narrative arc: from 'unexpected' and 'seems wrong' to 'absolutely brilliant.' This technique is designed to heighten anticipation and make the eventual reveal more impactful, rather than providing objective information. It's a rhetorical device to engage the audience, not a factual claim.
Original quote: “As always, not only have they open- sourced the model, but they've also released a technical paper on this. And how they designed this is just really unexpected and sometimes even seems wrong. But once you understand the logic behind everything, then it suddenly becomes absolutely brilliant.”
Paints Deepseek as an underdog with 'dozens of times smaller' team and no good GPUs — sets up an emotional button. — Emotional Button (45/100)
At 0:47
The 'small Chinese lab' with 'dozens of times smaller' team and no 'best Nvidia GPUs' is a classic underdog story — designed to make you root for them. 😈
Why this score: The speaker is using loaded language to create an emotional connection, emphasizing Deepseek's perceived disadvantages (small lab, less funding, smaller team, no top GPUs, no massive data center) to amplify the impact of their later achievements. It's a setup for the 'impossible' feat. While these constraints might be true, the framing is clearly for emotional effect.
Original quote: “Keep in mind that this is just a small Chinese lab that does not have nearly as much funding as OpenAI. Their team is like dozens of times smaller. Plus, they don't even have access to the best Nvidia GPUs out there, nor do they have a massive data center. In fact, they are severely constrained in…”
Claims Deepseek V4.1 Flash 'matches the performance of Frontier models' despite being a 'Flash model' — a confidence mismatch. — Confidence Mismatch (45/100)
At 1:06
Matching 'Frontier models' is a bold claim, especially for a 'Flash model.' Where's the data, mortal? 💀
Why this score: The speaker makes a strong claim about Deepseek V4.1 Flash matching 'Frontier models' without providing any specific benchmarks, comparative data, or a definition of what 'Frontier models' entails in this context. The confidence in this performance claim, especially given the previous 'underdog' setup, feels unearned without supporting evidence. It's a big statement with no immediate receipts.
Original quote: “Yet, they just released their latest model, Deepseek V4.1 Flash, which even matches the performance of Frontier models, even though this is just a Flash model.”
States memory footprint is 'over 400 times smaller' than the first generation — a specific, impressive stat. — No Frame (75/100)
At 1:19
A specific, massive improvement. This is the kind of number that actually means something. 🔥
Why this score: The speaker provides a very specific and quantifiable claim about the memory footprint reduction ('over 400 times smaller'). This is a clear, measurable achievement that, if true, represents significant progress. It's presented as a direct fact without obvious rhetorical manipulation, unlike some of the broader performance claims.
Original quote: “And get this, its memory footprint is like over 400 times smaller compared to the first generation.”
Explaining KV cache with a student-taking-notes analogy — a straightforward comparison. — No Frame (75/100)
At 2:49
A simple analogy to explain a technical concept. Nothing tricky here, just setting the stage. 😈
Why this score: The speaker uses a relatable analogy of a student taking notes to explain the function of a KV cache in AI models. This is a clear, unembellished explanation designed to simplify a complex topic for the audience.
Original quote: “To understand this a bit better, here's a nice analogy. Think of an AI model like a student watching a really long lecture and then taking notes along the way. Later, when the student needs to answer a question about the lecture, he doesn't need to replay the entire lecture from the beginning.…”
Framing the 'problem' as the industry's focus on long, complex AI tasks — setting up a challenge. — No Frame (75/100)
At 3:17
He's laying out the current challenge in AI development. It's a clear statement of industry direction. 😈
Why this score: The speaker describes the current industry trend towards developing AI models capable of handling longer, more complex tasks over extended periods. This is presented as a factual observation of the direction of AI research and development, not a manipulative claim.
Original quote: “Now, here's the problem the industry is facing right now. You see, for a short prompt like this, everything works fine. The AI model can easily convert this into a fairly small KV cache that fits well within its active memory. But here's the thing, the industry is now focused on making AI handle…”
Extending the student analogy to 'massive' notes filling a room — an emotional exaggeration. — Emotional Button (45/100)
At 3:47
He's taking the analogy to an absurd extreme to make a point. 'Notes filling a room' is a bit much, isn't it? 💀
Why this score: The speaker extends the student-notes analogy to an exaggerated degree, describing notes 'filling up his entire room' and needing 'filing cabinets in another room down the hallway.' While the core concept of a growing KV cache is valid, this dramatic imagery is designed to evoke a sense of overwhelming scale and difficulty, rather than providing a precise technical description.
Original quote: “And in these scenarios, the KV cache or the notes that the AI has to take is going to be massive. So going back to our analogy now, instead of just watching one lecture, the student has to watch weeks and weeks of online lectures and take notes on all of them. His notes will start to pile up fast…”
Explaining HBM and its limitations with a desk analogy — straightforward setup. — No Frame (75/100)
At 4:30
He's just laying out the technical basics here. No tricks, just setting the stage for the problem. 😈
Why this score: The speaker is using a simple, relatable analogy (student's desk) to explain high bandwidth memory (HBM) and its limitations (limited, expensive space). This is a clear, unembellished explanation of a technical concept, not a manipulative frame.
Original quote: “information. You see, inside a GPU, you have this high bandwidth memory or HBM. This is the incredibly fast memory that's right next to the processor chip. It's really fast and easy to access. So, it's kind of like the surface of a student's desk in our analogy. any notes that are sitting on this…”
Comparing SSDs to filing cabinets, calling latency 'devastating' — a bit dramatic. — Loaded Language (45/100)
At 4:57
He's right about the latency, but 'devastating'? That's a bit much for a technical bottleneck. It's an inconvenience, not an apocalypse. 💀
Why this score: The speaker accurately describes the technical trade-off between HBM and SSDs (more space, but higher latency). However, using the word 'devastating' to describe the latency, while technically true that it's a significant performance hit, adds an unnecessary emotional weight to a purely technical explanation. It's loaded language to emphasize the problem.
Original quote: “So the computer has to start storing this data somewhere else. For example, on solidstate drives or SSDs which are outside the GPU. So, in our analogy, it's like putting some of the students notes into filing cabinets in another room down the hallway. You have much more storage space there, but…”
Describing the data transfer as the 'absolute bottleneck' — accurate but still a bit hyperbolic. — No Frame (75/100)
At 5:36
He's not wrong. When the processor waits, that's the bottleneck. It's just how these things work. 😈
Why this score: The speaker's description of data transfer speed becoming the 'absolute bottleneck' when the processor is idle waiting for data is a technically accurate assessment of the problem. While 'absolute' might be strong, it correctly identifies the core issue in this scenario without resorting to manipulative language or misrepresentation.
Original quote: “If the KV cache needs to be stored there, well, every time the AI needs to predict the next word, it has to fetch data from the SSD, pull it through the motherboard up into its high bandwidth memory, and then into the GPU processor. This data transfer speed becomes the absolute bottleneck. The…”
Summarizing two 'major problems' as compute and speed, using the student analogy — clear, but simplifying. — No Frame (75/100)
At 5:58
He's just recapping the two main issues he's been explaining. It's a summary, not a trick. 😈
Why this score: The speaker is summarizing the two main technical challenges (compute and speed) using the established student/notes/filing cabinet analogy. This is a straightforward recap of the information already presented, intended to clarify the problem statement before introducing the solution. No manipulative framing detected.
Original quote: “So, to sum things up, there are two major problems that current AI systems are facing. One problem is compute. The student is basically just drowning in his own notes and it's really hard for him to find things. The second problem is speed. Because there are so many notes, some of these notes are…”
Explaining the KV cache generation — straightforward technical detail. — No Frame (75/100)
At 7:04
Just laying out the basics of how these models work. No tricks here, just the setup. 😈
Why this score: The speaker is explaining a fundamental component of transformer architecture, the KV cache, which is a standard concept in large language models. This is foundational information for the subsequent explanation of Deepseek's innovation.
Original quote: “Now, each layer that it goes through generates its own notes or KV cache.”
Introducing Deepseek's split architecture — highlighting its novelty. — No Frame (75/100)
At 7:19
Describing Deepseek's core innovation — a genuinely different approach. No hyperbole, just the facts. 😈
Why this score: The speaker is introducing Deepseek's novel architecture, specifically the split into causal encoder and decoder components. This is a direct description of their design, which indeed deviates from standard transformer models, making it a straightforward factual claim.
Original quote: “Well, the architecture from this new deepse has many layers. But here's the really unusual part. They split this into two halves. There's a causal encoder component and then there's a decoder component. This already looks completely different from a standard transformer model.”
Simplifying the decoder's inactive role during prefill — a bit too casual. — Loaded Language (45/100)
At 7:41
Calling it 'doing nothing' is a bit dramatic. It's inactive, not just sitting there picking its nose. 💀
Why this score: While the decoder component is indeed inactive during the prefill phase, describing it as 'doing nothing' is an oversimplification. It's not 'doing nothing' in a colloquial sense, but rather its specific computational role is not engaged at that stage. This phrasing adds a touch of casual exaggeration.
Original quote: “The last half is essentially turned off. They just do nothing.”
Explaining the 'genius' of borrowing the KV cache — clear and concise. — No Frame (75/100)
At 8:12
Explaining the core cleverness of Deepseek's design. It's a smart move, no argument there. 🔥
Why this score: The speaker is detailing the key innovation of Deepseek's architecture: the decoder's ability to 'borrow' the global KV cache from the encoder. This is a direct explanation of how their split design addresses the memory challenge, and the term 'genius' here is a subjective but understandable descriptor for an effective technical solution.
Original quote: “Here's why this design is so genius. Instead of having to read everything itself, it just looks at the output from the last layer of this encoder component. In other words, it borrows the completed global KV cache directly from the end of this encoder block.”
Explaining a complex AI architecture — no tricks, just information. — No Frame (75/100)
At 8:30
He's laying out the technical details of Deepseek's design. Straightforward explanation, for once. 😈
Why this score: The speaker is describing the specific architectural design of Deepseek's AI model, differentiating between global and local context processing. This is a technical explanation, not a persuasive claim or a rhetorical trick.
Original quote: “block. Now, this is quite a shocking and unexpected design because if they did this, if half its brain basically skipped the reading part, doesn't it lose some understanding of the context, which would make its answer worse? Well, that's the exact risk of this architecture. And that's why it took…”
Defining 'global' vs. 'local' context with an analogy — clear and direct. — No Frame (75/100)
At 8:58
He's clarifying technical terms with a simple analogy. It's a rare moment of clarity. 😈
Why this score: The speaker is defining 'global context' and 'local context' in the realm of AI processing, using a student's lecture notes as an analogy. This is a helpful clarification, not a manipulative framing technique.
Original quote: “Now, you might be wondering, what's the difference between global and local context here? Well, global context is basically all the information that was given to it plus your prompt and everything else attached. It's basically all the notes that the student took after watching weeks and weeks of…”
Describing the 'sliding window attention' mechanism — technical detail. — No Frame (75/100)
At 9:24
More technical explanation of how the AI processes information. Still just explaining the tech. 😈
Why this score: The speaker is detailing how the decoder component of the AI model uses 'sliding window attention' to focus on recent tokens rather than the entire global context. This is a technical description of a feature, not a rhetorical device.
Original quote: “You'll see this decoder component doesn't calculate any global KV cache, but instead it uses something called a sliding window attention to pay extremely close attention to the most recent tokens only, but not everything before it.”
Using a 'junior analyst/senior executive' analogy to explain AI processing — clear illustration. — No Frame (75/100)
At 9:40
He's using a corporate analogy to make the AI's processing easier to grasp. It's a decent illustration, I'll give him that. 😈
Why this score: The speaker employs an analogy of junior analysts creating an executive summary for senior executives to explain how the AI model processes information, with the 'junior analysts' being the first half of the model and 'senior executives' being the decoder layers. This is an illustrative technique, not a deceptive one.
Original quote: “And it turns out that this design works pretty well. Here's a nice analogy to wrap your head around this. Imagine a company where they need to analyze thousands of pages of financial reports. Well, they would first get junior analysts to read every single page and crunch out the numbers, do the…”
DeepC bypasses half the model, slashing compute by half — a bold claim of efficiency. — Confidence Mismatch (45/100)
At 10:30
He says 'slashes compute by half' like it's a done deal, but the 'how' is still a question mark. Bold claims, thin details. 💀
Why this score: The speaker confidently states DeepC 'completely bypasses' half the model and 'slashes compute required to read things by half' as an established fact. However, he immediately pivots to 'we still have the problem of memory' and 'how on earth can we reduce these massive piles of notes?', implying the solution isn't fully realized or explained yet. It's a classic move: declare victory, then admit the battle's still raging.
Original quote: “scrutinize the exact wording of the page sitting right in front of them. That's basically the local context. The senior executives rely on this global summary for direction, but they mostly focus locally on the page in front of them for execution. And by structuring the model this way, DeepC…”
DeepSeek V4.1 Flash shrunk KV cache size by 437 times compared to V1 — a dramatic comparison. — No Frame (75/100)
At 13:01
A 437x reduction from V1 to V4.1 Flash is a hell of a leap. The numbers are right there. 🔥
Why this score: The speaker provides specific, verifiable numbers for the KV cache size per token for DeepSeek V1 (390,000 bytes) and V4.1 Flash (890 bytes). The calculation of a 437x reduction is mathematically sound based on these figures. This is a direct, data-driven comparison presented without obvious manipulation.
Original quote: “If you compare the global KV cache size per token, which is basically the size of the notes the student has to take, DeepSeek V1 is almost 390,000 bytes. Now, if you fast forward just a few generations to this latest V4.1 Flash, it's only 890 bytes per token. They basically shrunk the size of the…”
DeepSeek V4.1 Flash is almost four times smaller than the previous V4 Flash — a direct, verifiable comparison. — No Frame (75/100)
At 13:24
Comparing 890 bytes to 3,500 bytes? Yeah, that's roughly four times smaller. The math holds up. 😈
Why this score: The speaker provides a specific comparison between DeepSeek V4.1 Flash (890 bytes per token) and the previous DeepSeek V4 Flash (3,500 bytes per token). The claim that the new update is 'almost four times smaller' is accurate (3500 / 890 ≈ 3.93). This is a clear, factual comparison.
Original quote: “Even if you compare this to the previous DeepSseek V4 Flash, this one still required like 3,500 bytes per token. So, this new update is like almost four times smaller than the previous generation. But here's the challenge to all of this. How can you compress these notes so much without losing the…”
DeepSeek's CSA2 mechanism shatters redundancy with 'extreme sharing' and three modes. — No Frame (75/100)
At 14:30
He's laying out the core concept of DeepSeek's new tech — explaining the mechanism, not making a grand claim yet. Just setting the stage. 😈
Why this score: The speaker is introducing a technical concept (CSA2 mechanism, extreme sharing, operating modes) and then immediately explaining how each mode works. This is foundational information for the rest of his argument, presented clearly without rhetorical tricks.
Original quote: “this new CSA2 mechanism by DeepSeek completely shatters this redundancy. It introduces the concept of extreme sharing. So instead of creating new notes from scratch every time, each layer could use three different operating modes. Full mode, reindex mode, and reuse. Let's go over each one. So if…”
Reindex mode reuses notes but creates a new index for different search paths. — No Frame (75/100)
At 15:15
Still explaining the tech, giving a clear example of how 'reindex' works. No spin, just information. 😈
Why this score: The speaker continues to describe the technical details of the 'reindex' mode, using a clear analogy (history of the world, chronological vs. thematic indexes) to illustrate its function. This is a straightforward explanation of a feature, not a claim requiring external verification.
Original quote: “Another mode is called reindex. And here's where the efficiency kicks in. If the layer is in reindex mode, it just reuses the notes or in other words the KV cache from a layer in full mode. It doesn't create its own notes from scratch. But what it does do is create a new index from scratch. Again,…”
Reuse mode achieves 'maximum efficiency' with 'almost zero memory effort'. — Loaded Language (45/100)
At 16:08
He says 'maximum efficiency' and 'almost zero memory effort' like it's a given. That's a bold claim without any numbers. 💀
Why this score: While describing the 'reuse' mode, the speaker uses strong, qualitative terms like 'maximum efficiency' and 'almost zero memory effort'. These are subjective descriptors that lack specific, quantifiable metrics or comparative data to back them up. It's an assertion of performance rather than a neutral description of function, using emotionally positive language to frame the feature.
Original quote: “And then finally, we have the third mode, which is maximum efficiency. And this is the reuse mode. If the layer has this mode, it exerts almost zero memory effort. It just reuses the notes from the full layer as well as the indices from either the full layer or the reindexed layers. It doesn't…”
Setting up a straw man to knock down — a classic move. — Straw Man (20/100)
At 17:36
He's building a hypothetical problem just to show how Deepseek 'solves' it. Predictable. 😈
Why this score: The speaker introduces a potential negative consequence ('it might hallucinate or miss important details') as an obvious expectation, creating a straw man argument. This sets up Deepseek's solution as more impressive by first presenting a common, easily refutable counter-argument that the technology supposedly overcomes.
Original quote: “Now, obviously, as you may expect, if we”
Deepseek's 'genius' in making it work despite restrictions — a confidence mismatch. — Confidence Mismatch (45/100)
At 17:48
He's calling it 'genius' and 'profound elegance' without showing the actual proof. That's just hype, not data. 💀
Why this score: The speaker uses highly laudatory language ('genius behind this,' 'elegance of their engineering is just profound') to describe Deepseek's achievement without providing specific, verifiable metrics or a direct comparison to other models that would justify such strong claims. It's a statement of belief, not a presentation of evidence.
Original quote: “everything. But here's the genius behind this. Deepseek was able to make it work”
Deepseek's resource limitations as a narrative device — emotional button. — Emotional Button (45/100)
At 18:06
He's playing the underdog card, making you root for them because they're 'starved for compute.' It's a classic emotional appeal. 🔥
Why this score: The speaker emphasizes Deepseek's lack of top-tier resources ('don't have the best Nvidia GPUs,' 'don't have the biggest data center,' 'starved for compute'). This narrative aims to evoke admiration and sympathy, making their achievements seem more impressive due to perceived adversity, rather than focusing solely on the technical merits.
Original quote: “They don't have the best Nvidia GPUs. They don't have the biggest data center in the world. Heck, they're pretty starved for compute.”
Analogy for AI memory — sets the stage for the problem. — No Frame (75/100)
At 18:30
He's just setting up the problem with a relatable analogy. No tricks here, just explaining the tech. 😈
Why this score: The speaker uses a 'senior executive' analogy to explain the concept of 'sliding window attention' as the AI's short-term memory. This is a straightforward explanation, not a manipulative framing technique. It's a setup for the problem DeepSeek solved.
Original quote: “This is where the senior executive doesn't read everything, but when he signs off on something, he has to look really closely at all the information on the page in front of him. Well, this sliding window attention is the AI's hyper local short-term memory.”
AI's short-term memory 'clogging hard drives' — problem statement. — No Frame (75/100)
At 18:43
Describing the technical challenge of AI memory storage. It's a clear problem statement, not a trick. 😈
Why this score: The speaker is explaining a technical challenge: the AI's short-term memory (local context) needs to be saved to SSDs to maintain conversation flow, but this process was 'clogging the hard drives.' This is a factual description of a known issue in AI memory management, as stated 'according to the paper.' No framing detected.
Original quote: “And according to the paper, this hyper local memory was actually causing a huge storage problem. You see, when a user has a conversation with the AI over multiple turns. In other words, when you say something, the AI replies and then you reply back and so on. The AI has to save this local context…”
Analogy for slow data retrieval — reinforces the problem. — No Frame (75/100)
At 19:12
Another analogy to make the problem clear. It's just good explanation, not manipulation. 😈
Why this score: The speaker uses a 'cabinets in another room' analogy to illustrate why fetching large amounts of cached data becomes slow. This is a continuation of the explanatory narrative, making a complex technical issue more accessible. It's a clear, non-manipulative way to convey the challenge.
Original quote: “Going back to our analogy, this is like filling up all the cabinets in the other room down the hall. Now, when you cache data, it means you save it in a temporary location so you can retrieve it quickly later, right? But if it gets too big and these notes are located in cabinets in another room,…”
Recapping previous points to set up a new problem — a straightforward summary. — No Frame (75/100)
At 20:30
Just laying out the groundwork for the next 'impossible' feat. Standard setup, no tricks here.
Why this score: The speaker is simply reiterating previously explained concepts to establish the context for the new, seemingly contradictory solution DeepSeek proposes. It's a clear, logical progression of the argument.
Original quote: “I just spent the past few minutes in this video explaining how they tried to reduce the compute and split the brain into halves to avoid generating and reading so many notes.”
Posing a rhetorical question to highlight a perceived inefficiency. — No Frame (75/100)
At 20:38
Setting up the problem they're about to solve. It's a classic narrative hook, nothing devious. 😈
Why this score: The speaker is using a rhetorical question to anticipate a listener's potential objection or confusion. This technique builds suspense and prepares the audience for the 'masterclass in efficiency' that DeepSeek will supposedly provide. It's a common storytelling device, not a manipulative frame.
Original quote: “But now, if we get this AI to recalculate this last part of the conversation every time, wouldn't this be incredibly inefficient?”
Describing the 'normal' slow method of data transfer — a clear explanation. — No Frame (75/100)
At 20:57
Explaining the standard, inefficient process. It's a clear, technical description. No hidden agenda here.
Why this score: The speaker is providing a straightforward technical explanation of a common data handling method, highlighting its inherent slowness due to physical data movement. This serves as a baseline for comparison with DeepSeek's innovative approach.
Original quote: “The first way is the normal way where we take data from the GPU, we push it through the motherboard and write it into a hard drive. And later, when it needs to use this shorter memory, it needs to search for it, pull it back up through the motherboard and into the GPU processor. This is super slow…”
Claiming GPU recalculation of 128 tokens is a 'microsecond operation' — a specific, verifiable technical claim. — No Frame (75/100)
At 21:16
A modern GPU can indeed crunch small token sets incredibly fast. This isn't hyperbole, it's just how powerful they are. 🔥
Why this score: The speaker is making a specific technical claim about the speed of modern GPUs for a particular task (recalculating 128 tokens). While 'microsecond operation' is a strong claim, the general principle that GPUs are extremely fast for parallel computation, especially on small data sets, is well-established in computer science. This isn't a framing trick; it's a technical assertion that can be evaluated.
Original quote: “Now, the second way is to just use the raw power of a modern GPU to simply recalculate its short-term memory. In other words, just rewrite the most recent and relevant notes from scratch. And for a modern GPU, just crunching like 128 tokens is just a microscond operation that barely takes up any…”
Attributing DeepSeek's innovation to realizing hard drive storage was a 'massive waste' — a clear statement of their insight. — No Frame (75/100)
At 21:41
This is the core insight they're presenting. It's a direct explanation of DeepSeek's strategy, not a rhetorical flourish.
Why this score: The speaker is explaining DeepSeek's rationale for their approach: identifying hard drive storage as an inefficient use of resources for short-term memory. This is presented as a direct observation and a key driver behind their architectural choice, framed as a logical conclusion rather than a manipulative statement.
Original quote: “The Deep Seek team realized that saving this short-term memory to the hard drive was just a massive waste of a really slow resource. By completely removing the step and just getting the GPU to quickly recalculate everything on the fly, not only does it make things faster, but it also freed up a ton…”
Summarizing DeepSeek's various optimization techniques — a straightforward recap. — No Frame (75/100)
At 22:00
Just a recap of the technical points already made. No new claims, just reinforcing the narrative. 😈
Why this score: The speaker is providing a concise summary of the different optimization strategies attributed to DeepSeek, as previously discussed in the video. This is a standard review technique to ensure the audience retains the key information before moving on. It's a factual recap, not a framing device.
Original quote: “Let's review what we've gone over so far. Deepseek used this sliding window attention to focus on the most immediate bits of information. They also split the model in half and then used hierarchical sparse indexing to significantly reduce the number of notes that the final layers need to process.…”
Explaining Single Pass MHC and its role in reducing data transfer bottlenecks. No tricks here. — No Frame (75/100)
At 22:30
He's laying out the technical problem and Deepseek's solution clearly. No smoke and mirrors, just the facts.
Why this score: The speaker is explaining a technical concept (Single Pass MHC) and its function in optimizing AI model performance by reducing data transfer latency. The explanation is straightforward, without loaded language or attempts to mislead. It's a clear description of a proposed technical improvement.
Original quote: “to save hard drive space. But guess what? We're not done yet. So DeepS also added some additional components that support the main architecture. One of them is called the single pass MHC. Now, to understand why this matters or what this is, you first need to know that running an AI model isn't just…”
Introducing the 'engram' with 168 billion parameters for static facts. Bold claim, but it's a technical description. — No Frame (75/100)
At 23:20
He's describing a specific architectural component and its parameter count. It's a technical detail, not a sales pitch.
Why this score: The speaker is detailing the 'engram' component, stating it has '168 billion parameters' and its purpose is to store 'static facts' in standard RAM, freeing up GPU memory for 'active reasoning'. This is a specific technical claim about a design choice and its intended benefits. While the impact is presented positively, the description itself is factual about the component's design and scale, not an exaggeration or misrepresentation of its nature.
Original quote: “Well, single path MHC is designed to eliminate some of these unnecessary trips. It's quite technical, but basically it mathematically aligns multiple operations together so that they can occur simultaneously. It's basically combining steps instead of running them one by one. And it turns out that…”
Comparing AI to a 'brilliant senior lawyer' — a loaded analogy. — Loaded Language (45/100)
At 24:44
He's dressing up the AI's function with a 'brilliant senior lawyer' analogy. It's a nice story, but it's just a story. 😈
Why this score: The speaker uses an emotionally positive and highly skilled human profession (brilliant senior lawyer) to describe the AI's function. This creates a positive association and elevates the perceived capability of the AI beyond a simple technical explanation, making it sound more sophisticated and intelligent than a direct description might.
Original quote: “It's kind of like a really brilliant senior lawyer working on a case and synthesizing information.”
Calling AI's function 'real strategic reasoning work' — a confidence mismatch. — Confidence Mismatch (45/100)
At 25:12
He's calling it 'real strategic reasoning' like it's a human brain. It's code, not consciousness. 💀
Why this score: The phrase 'real strategic reasoning work' attributes a high-level cognitive function, typically associated with human intelligence and consciousness, to an AI system. While AI can perform complex tasks, framing it as 'real strategic reasoning' overstates its current capabilities and implies a level of understanding and intent that isn't yet verifiable, especially without further technical detail.
Original quote: “focus on the real strategic reasoning work.”
Summarizing 'sliding window attention' as only focusing on 'immediate bits' — a simplification. — No Frame (75/100)
At 25:58
He's just summarizing a technical concept here, no real trick. It's a straightforward description. 😈
Why this score: The speaker is providing a high-level summary of 'sliding window attention,' which is a known technique in AI for managing context length by focusing on recent information. While simplified, it accurately describes the core function without manipulative language or omissions, acting as a recap rather than a new claim.
Original quote: “They've added this sliding window attention to only focus on the most immediate bits of information.”
Claiming 'completely deleted its short-term memory' — a bold, potentially misleading statement. — Loaded Language (45/100)
At 26:22
He says 'completely deleted' its memory like it's a lobotomy. It's a technical optimization, not an exorcism. 🔥
Why this score: The phrase 'completely deleted its short-term memory' is strong and absolute. While the 'bounded replay mechanism' likely optimizes memory usage by discarding less relevant information, 'completely deleted' might be an overstatement for dramatic effect. It implies a more drastic and perhaps less nuanced change than a technical explanation would convey, potentially misrepresenting the actual function for impact.
Original quote: “We also completely deleted its short-term memory with this bounded replay mechanism.”
Calling it 'Frankenstein of efficiency' then 'most optimized' — pure hype before the data. 😈 — Loaded Language (45/100)
At 26:49
He's piling on the superlatives before showing a single chart. That's not data, that's a sales pitch with a graph coming. 🔥
Why this score: The speaker uses highly charged, positive adjectives like 'Frankenstein of efficiency,' 'optimized,' 'turbocharged,' and 'frictionless' to build extreme anticipation and positive sentiment before presenting any evidence. This is classic emotional priming, making the audience more receptive to the data that follows, regardless of its actual impact.
Original quote: “They've built the most optimized, turbocharged, frictionless Frontier model we've seen so far. But don't take my word for it.”
Claims it 'defies the laws of AI' by flattening the curve — bold statement for an engineering feat. 💀 — Confidence Mismatch (45/100)
At 27:47
Calling an engineering breakthrough 'defying the laws of AI' is a bit much. It's an innovation, not a miracle. 🙄
Why this score: The phrase 'completely defies the laws of AI' is an overstatement. While Deepseek's achievement of a flat decode curve for increased context window size is a significant engineering advancement, it doesn't 'defy' fundamental laws. It represents a clever solution or optimization within existing computational principles, not a magical breaking of them. This hyperbolic language inflates the significance beyond what the technical description implies.
Original quote: “This is a massive deal. It completely defies the laws of AI because normally what we would expect is if you scale the context, the compute also increases as you could see with the previous generations of deepseek models. But here they've basically flattened the curve.”
Deepseek 1.1 scores 74.2, 'on par' with GPT6 Astra's 74 — a tiny difference framed as equal. 😈 — Loaded Language (45/100)
At 28:01
He says 'on par' for a 0.2 difference. That's not 'on par,' that's 'slightly better' — but he wants you to feel the equality. 🔥
Why this score: The speaker claims Deepseek 1.1's score of 74.2 is 'on par' with GPT6 Astra's 74. While numerically close, framing a slight lead as 'on par' downplays Deepseek's actual (albeit small) advantage, aiming to emphasize its competitive standing against a well-known 'Frontier model' rather than its precise ranking. It's a subtle manipulation of perception to elevate Deepseek's status.
Original quote: “And you know the ridiculous thing is not only is this super efficient and frictionless, but its performance is also state-of-the-art, pretty much on par with the Frontier models. For example, if you look at Deep Su 1.1, this scores 74.2, which not only beats the other open models out there, but if…”
Deepseek beats GPT6 Astro Max and is #1 on LiveBench. Bold claims, but where's the chart? — Missing Context (45/100)
At 28:31
He's rattling off rankings and beating 'GPT6 Astro Max' like it's gospel — but the actual leaderboard isn't on screen. Show the damn proof. 💀
Why this score: The speaker makes a definitive claim about Deepseek's performance against a specific, unreleased model (GPT6 Astro Max) and its ranking on 'LiveBench' without displaying the referenced leaderboard or chart. It's all talk, no visual evidence to back the specific numbers or rankings. I've seen this trick since before your grandfather tried it. 🔥
Original quote: “You can see that this freak of a model even beats GPT6 Astro Max. If you look at this leaderboard by LiveBench, then you can see that this new Deepseek is the number one ranked open model.”
Deepseek 4.1 Flash is #1 on Val's index. Another top ranking, another missing visual. — Missing Context (45/100)
At 28:42
Another 'number one' claim, this time for Val's index — still no visual proof. Just trust him, I guess. 😈
Why this score: The speaker asserts Deepseek 4.1 Flash is ranked number one on 'Val's index' for knowledge work tasks. Again, this is a specific, verifiable claim that would be easily proven by showing the index on screen. Instead, it's just stated as fact, leaving the viewer to take his word for it. It's a classic move: tell them it's true, don't show them why. 🚩
Original quote: “Same with Val's index, which measures an AI's performance across knowledge work tasks. You can see that Deepseek 4.1 Flash is currently ranked number one.”
Deepseek is the cheapest, with competitors 20x more expensive. Big numbers, but where's the cost breakdown? — Missing Context (45/100)
At 28:52
He's claiming 'insane cost' and '20 times more expensive' without showing the actual cost chart. That's not data, that's just confidence. 💀
Why this score: The speaker claims Deepseek is the cheapest and 'over 20 times more expensive' than second and third place. While a chart showing 'cost per task' is mentioned and presumably shown later, at this exact moment, the specific numbers or the chart itself are not visible to back up the '20 times' claim. It's a big number dropped without immediate visual support, relying on the listener's assumption. I've seen this trick a million times. 🔥
Original quote: “And look at the insane cost of this. Not only is this the most performant model, but it's also the cheapest. You can see that second and third place are over 20 times more expensive.”
Deepseek achieves 200 tokens/second, 4x faster than GPT6. Specific numbers, but still no visual proof. — Missing Context (45/100)
At 29:14
He's throwing out '200 tokens per second' and 'four times faster than GPT6' — specific, impressive numbers, but the chart isn't on screen yet. Just tell me what to believe, why don't you? 😈
Why this score: The speaker states Deepseek achieves 'over 200 tokens per second' and is 'four times faster than GPT6.' These are precise, quantifiable claims that demand visual evidence. While he might show a chart later, at this point, it's just a spoken statistic without the accompanying data visualization, making it an assertion rather than a demonstrated fact. It's the classic 'trust me, bro' with a technical twist. 🚩
Original quote: “If you look at the output speed if you use it through their API, again, this is insane. This achieves over 200 tokens per second, which is like four times faster than GPT6.”
Deepseek open-sourced the model, 'pulling off the impossible.' A claim of generosity and a grand feat. — No Frame (75/100)
At 29:43
He's just stating they open-sourced it and praising the team. No tricks here, just a pat on the back. 😈
Why this score: The speaker states that Deepseek has open-sourced its model, allowing users to download it locally. This is a straightforward claim about the availability of the model and a compliment to the team. There's no hidden agenda or manipulative framing here; it's a direct statement of fact and opinion. Fine. Green pill, moving on. ✅
Original quote: “And as always, they've open sourced the model for you to download locally, so you can do whatever you want with it. My hat off to the Deep Seek team for pulling off the impossible.”
See the full analysis with timestamps →