THE DECODER

According to Bloomberg, Safe Superintelligence (SSI) is close to completing another funding round of more than $1 billion that would value the company at over $30 billion.

1/ Researchers from Metas FAIR, the University Gustave Eiffel and EHESS show that AI systems can develop an intuitive understanding of physical laws through self-supervised training on videos without the need for pre-programmed rules. 2/ The team used the Video Joint Embedding Predictive Architecture (V-JEPA), which makes predictions in an abstract representation space. In tests with the violation-of-expectation paradigm, V-JEPA showed better physics understanding than large language models. 3/ The results are part of Meta's research into the JEPA architecture, which is intended to offer an alternative to generative AI models. The aim is to create comprehensive world models that enable autonomous AI systems to gain a deeper understanding of their environment. 👇 Read more #GenerativeAI #JEPA #MetaAI #YannLeCun

1/ OpenAI CEO Sam Altman seems to be aiming for a return to open source, asking users on X for feedback on whether the next project should be an o3-mini level model or the best smartphone model. 2/ OpenAI had distanced itself from open source after a billion-dollar investment from Microsoft and the release of GPT-4, but recently admitted that this strategy was wrong - probably also in light of competition from open source models from companies like Meta, Deepseek, Alibaba, and Mistral. 3/ However, this is not a return to old values, but rather a recognition that the completely closed approach no longer works when the competition is catching up so quickly. 👇 Read more #ChatGPT #OpenSource #OpenAI

1/ Elon Musk's AI company xAI has introduced its new language model Grok 3, which is available in several variants and features special "reasoning" models for mathematical and scientific questions. 2/ In initial tests, Grok 3 achieves top scores on the benchmarking platform lmarena.ai and impresses with strong logical reasoning capabilities, but also shows weaknesses in the new DeepSearch search function, which sometimes delivers incorrect results. 3/ With the release of Grok 3, xAI is raising its prices significantly: the Premium+ subscription on X now costs $50 per month, and there is also a SuperGrok subscription for $30 per month. The service is initially available only in the US, with more features such as a voice mode in the pipeline. 👇 Read more #GenerativeAI #Grok #Grok3 #xAI

1/ Mistral AI has introduced Saba, a new language model designed specifically for the Middle East and South East Asia, incorporating the linguistic subtleties, cultural context, and regional knowledge unique to these areas. 2/ Despite having only 24 billion parameters, which is considerably smaller than many other models, Mistral AI claims that Saba delivers superior accuracy and speed at a reduced cost. 3/ Saba is also compatible with less powerful hardware and can achieve speeds exceeding 150 tokens per second, making it more accessible and efficient for a wider range of users.

1/ A Microsoft study involving 319 knowledge workers reveals that the use of generative AI (GenAI) can impact critical thinking skills. 2/ When using GenAI, users tend to prioritize verifying AI-generated results rather than independently gathering information, incorporate AI answers instead of solving problems on their own, and focus on monitoring AI performance rather than directly executing tasks themselves. 3/ The researchers cite an "irony of automation" effect, where AI's takeover of routine tasks may deprive humans of opportunities to exercise their judgment, potentially leading to a weakening of their own abilities over time due to this "cognitive relief."

Elon Musk announced on X that his company xAI will showcase its latest AI model Grok 3 in a live demonstration on Monday at 8 PM Pacific time (Tuesday, 5 AM CET).

1/ A study involving 830 participants found that they could only slightly distinguish between therapeutic responses generated by ChatGPT and those provided by human therapists. 2/ Surprisingly, the AI-generated responses were perceived as more empathetic. Researchers suggest this may be due to factors such as the length of the responses, a more positive tone, and the use of more nouns and adjectives. 3/ The study also revealed a degree of skepticism towards AI, as responses believed to be generated by AI were rated lower than those attributed to human therapists. The highest-rated responses were machine generated but attributed to humans.

1/ Researchers from Meta and Spain have successfully reconstructed typed sentences using only non-invasive brain recordings from MEG and EEG. An AI model was able to learn and reconstruct the sentences with a character-level accuracy of up to 80 percent. 2/ The researchers also explored how the brain progressively converts thoughts into specific finger movements. They discovered that words and actions are represented in a coherent and simultaneous manner using a unique "dynamic neuronal code". 3/ Decoding this neural code of language is a significant challenge for both AI and neuroscience, particularly for the development of neuroprostheses that transmit signals to AI decoders.

Meta has begun developing both hardware and software for AI-powered humanoid robots, according to a new report from Bloomberg.