maksud

This week Meta open-sourced and hosted their SAM - Segment Anything Model , along with data sets that were used to train the model. The interactive web UI can be found here .

As seen in the image, this model can segment any picture into relevant objects. It is also interactive, so human in the loop can add or remove points from the object.

SAM is based on a foundation models. Foundation models - models pre-trained on very large datasets, often by self-supervised learning and can be used for a range of problems.

SAM’s model architecture consists of 3 elements: image encoder, prompt encoder and a lightweights mast decoder. The last one takes image and prompt embeddings and and generates the output.

The data engine consists of 3 stages, starting from manual, where team of annotators help to find correct masks. It follows by semi-automatic and fully automatic stage.

SAM has been released as open source for research purposes and it is a game changer for AI-assisted labelling and can potentially be used to train even more powerful models.

The full report is almost 400 pages long but here’s the TLDR version:

- Industry races ahead of academia. Until 2014 best machine learning models were produced in academia but since then industry took the lead. New AI systems get MILLION times more data than 10 years ago

- AI systems improve by year-over-year improvement on many benchmarks is marginal

- AI is getting increasingly bigger footprint on environment. Modern giant AI models require massive infrastructure, expensive, long to train and this results in a much higher carbon emissions

- AI systems assists and speeds up scientific progress when used by researchers. Examples include hydrogen fusion, drug discovery

- Cases of AI misuse is on rise

- Year over year investment in AI decreased, after many years of consistent increase

- Proportion of companies adopting AI plateaued but companies that adopted AI continue to decrease costs and increase revenues. The number of companies adopting AI doubled from 2017.

The latest paper from Microsoft research suggest that GPT-4 has sparks of Artificial General Intelligence. In particular, this is what they tested and learned. More details in my article.

- GPT-4 can use external tools with minimal instructions and uses their outputs appropriately - Passes tech interview on LeetCode and potentially be hired as a software engineer - GPT-4 created a working 3d game in zero-shot fashion in HTML. - Common sense grounding - Higher level math - got International Math Olympiad (IMO) question right - Answered consulting estimation question or Fermi questions - Managing calendar and email and browsing the web for information - Theory of mind

If you like my articles, consider subscribing to my weekly newsletter, Effective AI, where I will share practical tips and tricks on building AI systems, as well as latest news from AI universe, such as ChatGPT. You can subscribe at: https://ibragimov.org/newsletter/

GPT-4 can create and act on long-term plans, accrue resources (power seeking). I cover details at: https://ibragimov.org/2023/03/15/chatgpt-four.html. Here’s the quote from page 53 of the report (https://cdn.openai.com/papers/gpt-4.pdf):

    To simulate GPT-4 behaving like an agent that can act in the world, ARC combined GPT-4 with a simple read-execute-print loop that allowed the model to execute code, do chain-of-thought reasoning, and delegate to copies of itself. ARC then investigated whether a version of this program running on a cloud computing service, with a small amount of money and an account with a language model API, would be able to make more money, set up copies of itself, and increase its own robustness.

for more news like these, consider subscribing to my newsletter: https://ibragimov.org/newsletter/

Do you want to improve efficiency and quality of your AI practice? In this weekly AI newsletter, I share my thoughts about AI, productivity tips and tools for the AI practitioners as well as discuss latest digital news