Algo huele raro en Multiverse Computing

Value of the Information & Strength of the Argument

The value of the information is primarily in the critical dissection of public data. The hosts provide concrete details about the company’s investment, the Compactifai paper, and the Quasar announcement, and they expose logical inconsistencies, such as the naming of the 88% method and the fact that a standard quantization library outperforms it. The argumentation is sound and methodical: they first establish their baseline (state-of-the-art quantization), then compare the company’s numbers, then analyze the tokenizer test, and finally discuss the business model. They are careful to separate facts from inferences and explicitly state their subjective opinion. While some points rely on educated guesses (e.g., the expert pruning process), they are clearly labeled as such. The podcast excels at translating technical details for a general audience without oversimplifying, and it maintains a balanced tone by acknowledging the possibility that the company’s methods could be legitimate in the future.

Scientific Rigor, Source Quality, Title Accuracy

The analysis relies on publicly available sources: the company’s LinkedIn description, the Compactifai paper, the Quasar model announcement, and community discussions on Twitter. The hosts also reference the Artificial Analysis website for model comparisons, though no direct URL is provided. They explicitly note that they are not journalists and did not contact the company for a response, which is a potential limitation in rigor. The title ‘Algo huele raro en Multiverse Computing’ is a strong claim, but it is well-supported by the evidence presented. The podcast does not cite external scientific literature beyond the paper in question, but it does reference common technical concepts. Overall, the source quality is adequate for a critical opinion piece, though deeper validation of the technical claims would require access to the full paper and benchmark details.

The title accurately reflects the content: the hosts investigate and articulate multiple reasons to be skeptical about Multiverse Computing's claims and practices.

7/10

The analysis is based on publicly available data (company claims, a published paper, and the model release) and the hosts are transparent about their lack of internal information. They carefully scrutinize numbers and expose inconsistencies, such as the 88% and 93% compression variants being worse than standard 4-bit quantization in some metrics. However, some assumptions (e.g., about expert pruning) are speculative and not fully verified.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction to the topic: Multiverse Computing and the recent Quasar model announcement.
• 03:00Discussion of the company's background, investment (€500M), and vague LinkedIn description.
• 07:00Analysis of the Compactifai paper; comparison with standard 4-bit quantization, noting the 88% and 93% methods perform worse.
• 15:00Reaction to the Quasar 438B release: initial ambiguity about its origins, later revealed to be a compressed GLM 5.2.
• 20:00Examination of the tokenizer match between Quasar and GLM 5.2, confirming they share the same tokenizer.
• 25:00Discussion of the business model: whether the compressed model is cheaper to serve, and the implications of rebranding an existing model.

• La TERTULia de la Inteligencia Artificial - Blog — The official blog of the podcast, mentioned in the video description as a source for more information.