Learning Library

← Back to Library

December AI Surge: Robots, Gemini, Claude

Key Points

  • A new humanoid robot built by robotics firm Abtronic in partnership with Google DeepMind aims to give AI real‑world sensory data, which could help overcome the “pre‑training wall” and enable intelligence to scale beyond internet‑derived data.
  • Google released Gemini 2.0 “experimental thinking,” a model that outranked OpenAI’s GPT‑4 on leaderboards, delivering detailed critiques, rewrites, and human‑level intent explanations that make it useful for final‑draft content generation.
  • Claude (Anthropic) announced a long‑awaited Excel‑file understanding update, allowing the model to ingest and manipulate structured spreadsheets up to roughly 30 MB, addressing a major limitation for LLMs working with tabular data.
  • Rumors suggest OpenAI will unveil a new model dubbed “03” later today, following a pattern of frequent releases (e.g., the “12th day of Christmas” analogy) that may further intensify competition among the leading AI providers.
  • These four developments—robotics integration, Gemini 2.0’s performance, Claude’s Excel capability, and the upcoming OpenAI release—highlight a rapid escalation in both AI hardware and software capabilities as the industry races toward more versatile, real‑world‑aware intelligence.

Full Transcript

# December AI Surge: Robots, Gemini, Claude **Source:** [https://www.youtube.com/watch?v=izfgyzADwKo](https://www.youtube.com/watch?v=izfgyzADwKo) **Duration:** 00:05:54 ## Summary - A new humanoid robot built by robotics firm Abtronic in partnership with Google DeepMind aims to give AI real‑world sensory data, which could help overcome the “pre‑training wall” and enable intelligence to scale beyond internet‑derived data. - Google released Gemini 2.0 “experimental thinking,” a model that outranked OpenAI’s GPT‑4 on leaderboards, delivering detailed critiques, rewrites, and human‑level intent explanations that make it useful for final‑draft content generation. - Claude (Anthropic) announced a long‑awaited Excel‑file understanding update, allowing the model to ingest and manipulate structured spreadsheets up to roughly 30 MB, addressing a major limitation for LLMs working with tabular data. - Rumors suggest OpenAI will unveil a new model dubbed “03” later today, following a pattern of frequent releases (e.g., the “12th day of Christmas” analogy) that may further intensify competition among the leading AI providers. - These four developments—robotics integration, Gemini 2.0’s performance, Claude’s Excel capability, and the upcoming OpenAI release—highlight a rapid escalation in both AI hardware and software capabilities as the industry races toward more versatile, real‑world‑aware intelligence. ## Sections - [00:00:00](https://www.youtube.com/watch?v=izfgyzADwKo&t=0s) **Google’s New Humanoid Robot Initiative** - The speaker outlines Google DeepMind’s partnership with robotics startup Abtronic to launch a humanoid robot, stressing how real‑world sensory data from such bots could break the pre‑training ceiling and accelerate the scaling of AI intelligence. ## Full Transcript
0:00today December 20th is already a massive 0:02day in AI before open AI drops whatever 0:06they're dropping for the 12th day of 0:08Christmas and the rumor is it's 0:10something called 03 we will see but 0:13already we have four big developments 0:15for you number one abtronic which is a 0:18robotics company has collaborated with 0:20Google Deep Mind to release a humanoid 0:24robot powered by Google AI this is big 0:27from a long-term perspective because one 0:30of the significant bets on the ability 0:34of AI to continue to scale intelligence 0:37is that we find another big data pool so 0:40when Ilia suer talked about the internet 0:42being our biggest data pool and it's 0:44being something that is not renewable 0:46the only way around that is if you give 0:49AI access effectively to situational 0:53awareness of the real world Elon has 0:55called that out with his cars and his 0:57robots in this case Google's entering 0:59that Arena as well the idea is that when 1:03humans learn human babies take in far 1:06more data in their eyes and their ears 1:10than we give to even our largest large 1:12language models in their first three 1:14four years of life if you can give a 1:17robot that kind of data input from 1:20interacting with the real world as you 1:22would if it was a you know actual 1:24walking around robot well maybe that's a 1:27way to get through the pre-training wall 1:29and start to continue to scale 1:31intelligence so that's the Strategic 1:33reason why Google getting into the 1:35robotic space is so 1:36interesting but Google's not done yet so 1:40Google also in the last 24 hours 1:42released a model that took the top spot 1:44in the leaderboards from open AI 1:4701 it's called Gemini 2.0 experimental 1:52thinking it's available in the AI studio 1:55now from Google I have played with it it 1:58is amazing 2:00I gave it something that was a document 2:02that I thought was okay that uh Claude 2:05Sonet 3.5 had written and I said can you 2:08make this better it was a very short 2:10prompt I did not do my best job at 2:12structuring a prompt it came back with 2:15the most detailed critique of how to 2:17make the doc better rewrote the entire 2:19Doc and described it in a way I could 2:22understand and then gave me the human 2:23intent behind it like the reason why it 2:26did what it did in a way that a human 2:28could understand and it made a ton to 2:29sense 2:31I was shocked I ended up using Claude 2:33for 2:35formatting and that's not really what 2:37you're supposed to use a large language 2:38model for but this model was so good 2:41Gemini thinking was so good that I just 2:44didn't need to touch it like we've gone 2:46from like it can be a draft to this 2:48might be a final draft and that's a big 2:50step forward and again open a may drop 2:53something even better later today we 2:55will see so that's number two Gemini 2.0 2:58thinking check it out 3:00number three Claude released a long 3:03awaited update to excel understanding so 3:08Excel file understanding has been a huge 3:10issue for large language models Claude 3:11has been at the Forefront of using tool 3:13sets to understand these structured data 3:16sets and these tables and Claude 3:19released a update that 3:22essentially the anthropic team is going 3:24to let you handle a Excel file up to 30 3:28Megs in size larger than the normal 3:30context window and they're not really 3:32clear quite how they do it but at the 3:34end of the day Claud is going to be able 3:36to look across that entire spreadsheet 3:38and extract meaningful insights even if 3:40it exceeds the traditional definition of 3:42the context window and that may be as 3:44simple as they're adding a special 3:46context window that they can trigger 3:47when a very large Excel file goes in but 3:50it's still significant because 3:52structured data sets increase 3:54combinatorially in complexity the bigger 3:57they get and so a 30 megga Excel file 4:00like I've worked with those Excel files 4:02they're really really hard to understand 4:03for a human and so getting the ability 4:05to like pull that into an AI is a big 4:07step 4:08forward 4:10finally uh I did not know this until 4:13today but apparently AI is getting to 4:16the point where it can pass the mirror 4:18test so one of the classic tests for 4:21intelligence in the animal kingdom is 4:24can an animal recognize that an image in 4:27the mirror is itself my Corgi cannot do 4:31this my Corgi is dumb as a sack of 4:32hammers but there are animals that can 4:35do this gorillas can do this there are 4:37other animals that can do this as well 4:38it's a well-known test in biology and so 4:41of course people are wondering can AI do 4:43this and the answer is AI is getting 4:46better and better and better at this 4:48Claude has passed the self-awareness 4:50test now the mirror test if you take a 4:53screenshot of 4:54Claude and give it to Claude that's how 4:59you do the mirror test and Claude can 5:00pass that now it doesn't mean that 5:03Claude passes it as often as humans do 5:05uh they apparently have a benchmark for 5:07self-awareness I didn't even know this 5:09and AI uh the four class models score at 5:12about 50% on self-awareness and humans 5:14score above 90% I didn't know we didn't 5:17score 100% but I guess we don't uh maybe 5:20nothing is 100% in this world but 5:23anyway the point is that the four class 5:26models are significantly better than the 5:27three class models at self-awareness and 5:30we should expect them to continue to get 5:32better and yes that imposes really deep 5:35philosophical questions and we're going 5:37to be asking a lot more philosophical 5:39questions around AI in 2025 so that's 5:43your update I will drop something else 5:45later in the day uh as open AI has their 5:48final day release party but I thought 5:50this news was too important not to share