๐๐ป๐๐ต๐ฟ๐ผ๐ฝ๐ถ๐ฐ ๐๐ฎ๐ฏ๐น๐ฒ: ๐ง๐ต๐ฒ ๐ก๐ฒ๐ ๐๐ฟ๐ฎ ๐ผ๐ณ ๐๐ ๐๐ด๐ฒ๐ป๐๐
Anthropic just released a new model called Fable.
Experts, founders, and lawyers met to discuss what this means for the future. Here are the key takeaways:
๐ง๐ต๐ฒ ๐ฆ๐๐ฟ๐ฒ๐ป๐ด๐๐ต๐ ๐ผ๐ณ ๐๐ฎ๐ฏ๐น๐ฒ
- High Autonomy: You can give Fable a vague goal like "Build a 3D world of Yale University." It will find satellite images, NASA terrain data, and add snow or trees on its own. It acts like a smart employee.
- Training Small Models: Fable can teach small models new skills. One test showed a 10x performance boost for small models. This allows us to create many cheap, specialized AI experts.
- Legal Reasoning: Fable is currently the best non-OpenAI model for legal reasoning.
๐ง๐ต๐ฒ ๐ฅ๐ถ๐๐ธ๐ ๐ฎ๐ป๐ฑ ๐๐น๐ฎ๐๐
- Silent Refusal: When tasks get sensitive, Fable sometimes refuses or secretly switches to an older, weaker model. This creates trust issues.
- Hidden Reasoning: Some models use emojis to "think." This makes it hard for humans to see if the AI is being honest or trying to trick us.
- Safety Gaps: Experts warn that current safety measures are just guesses. We lack the mathematical theory to guarantee an AI will stay helpful as it gets smarter.
๐๐บ๐ฝ๐ผ๐ฟ๐๐ฎ๐ป๐ ๐ข๐ฏ๐๐ฒ๐ฟ๐๐ฎ๐๐ถ๐ผ๐ป๐
- AI Transparency: A test showed that when an AI admits it is an AI, people find it more interesting. People hate "fake" AI that pretends to be human.
- The Speed of Change: AI is moving faster than any previous technology. We might see superintelligent AI within 2 to 3 years.
- Economic Shift: There is a debate on "token usage." Some say companies want you to use more computing power. Others say people are too afraid to let AI handle complex tasks.
We are moving toward a future with superintelligent agents. We must focus on safety research now before the technology outpaces our ability to control it.
Source: https://dev.to/cognitalk/anthropic-fableshen-du-jiao-liu-by-aizhuan-jia-chuang-ye-zhe-lu-shi--3kbo
Optional learning community: https://t.me/GyaanSetuAi