Good morning, human brains. Welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

• Results from the largest-ever Turing Test 🔍️ 

  We’re close to convincing you that we’re humans.

• Humanoid showdown 🦿 

  A deep-dive into the 5 most promising humanoid robots.

• Xtalpi + Eli Lilly = drugs on AI 💉

  A $250M deal to create novel drugs.

Sponsor Bot Eat Brain | New here? Subscribe!

APPETIZER

Results from the world’s largest Turing test

AI21 Labs trialed the world’s largest Turing test, with the help of their game, “Human or Not?”. The gamified test captured 10 million conversations with more than 1.5 million participants since mid-April.

The Turing test was conceived by the famous British mathematician Alan Turing in 1950.

In it, a human evaluator engages in conversation with two participants: a human and a machine. The evaluator communicates with the two parties through a text interface. If the machine can successfully convince the evaluator that it is human more than 50% of the time, it passes the test.

AI21’s gamified version randomly matches human participants online with either participant or a chatbot. The chatbots included prompted versions of GPT-4, Jurassic-2 (AI21’s own chatbot), and Cohere. The prompts include fictitious locations, current weather, personas with backstories, and even current events.

Each participant has twenty seconds to type and send a message on their turn. After two minutes, the conversation ends and participants are asked whether their partner was human or AI.

The results?

Overall, 68% guess correctly — but there’s a fascinating discrepancy here. When talking to other humans, participants guessed correctly 73% of the time. When talking to the bots, however, participants guessed correctly only 60% of the time.

Uh oh.

AI21 offers several reasons for the asymmetry. People assume AI’s don’t make the same typos or use the same slang as humans (which these bots were trained to do) or they assume the AI’s are unaware of current events (in this case, they were prompted with a handy list of them).

Our take: Since there are artificial time limits in place, this is only a limited version of the Turing test. Still, the latest LLMs are already fooling you human brains 32% of the time, with just a little bit of prompt engineering.

How soon will LLMs pass this version of the Turing test? And then a full-fledged Turing test?

P.S. The game is still live, and we recommend you try it out. It’s as shocking as it is entertaining.

FROM OUR PARTNERS

Free DTC AI Training

• Tailored for DTC Brands

• Decrease your CAC by 30% in 72 Hours by saving 12 hours a week on reporting.

• Learn the simple 3-step process in just 3 videos. Calculator marketing strategy blueprint, ecommerce metrics maps, and relationships.

• These strategy sessions and roadmap normally sell for $2000 — we’re offering it for FREE.

BUZZWORD OF THE DAY

MAIN COURSE

The humanoid robot race 🏎️ 

Ever wish you had a robot butler?

Science fiction is quickly becoming a reality in the race to build the world’s first useful humanoid robot and startups have raised hundreds of millions of dollars to fund their development.

We researched the five most promising humanoid robots and their current capabilities. Here are the companies to watch:

1/ Prosper’s Alfie

Movement: Wheeled base
Torso: Vertical spine with independent arms and head
End effector type: Two-jawed gripper with customized tools
Current level of autonomy: Teleoperated by humans
Application: General purpose, tailored for household and office chores
Estimated max payload: Unspecified

Based out of London, Prosper Robotics is betting on a utilitarian approach to household and office robot helpers.

By our estimates, it’s the lowest-cost robot out of these five, with an estimated price of £5k-10k. Its grippers have limited dexterity, but a suite of tools should extend gripping capability.

While it won’t be able to climb stairs, its arms and head can move up and down its spine, giving it the widest vertical range out of the playing field here.

[Disclosure: The author of this article previously worked at Prosper]

2/ Tesla’s Optimus

Movement: Articulated legs
Torso: Hinging
End effector type: Five-fingered articulated hand with opposable thumbs
Current level of autonomy: Teleoperated/trained by humans
Application: General purpose
Estimated max payload: 20kg

One of the more complex designs, Optimus has been engineered entirely from the ground up. In classic Tesla fashion, it features custom actuators, which gives them more control over payload, form factor, and fine coordination.

Optimus also has an edge with fine object manipulation due to its five-fingered hand design.

How quickly Tesla will move towards full autonomy?

3/ Sanctuary AI’s Phoenix

Movement: Articulated legs with hips
Torso: A hinging torso
End effector type: Five-fingered articulated hand with opposable thumbs
Current level of autonomy: Teleoperated by humans
Application: General purpose work
Estimated max payload: 25kg

Similar to Tesla’s expensive and intricate approach, Phoenix features articulated legs, but with a different hip geometry — relying on each leg’s independent hip joint to balance instead of a central pivoting joint. Notably, Sanctuary has chosen a hands-centric approach.

Finalizing the hands is the hardest part, and nailing that down first guarantees more stability in the rest of the design.

4/ The Figure 01

Movement: Articulated legs with hip joint
Torso: Hinge at the hip
End effector type: Five-fingered articulated hand with opposable thumbs
Current level of autonomy: Unknown
Application: General purpose labor, household, and “off-world”
Estimated max payload: 20kg

Figure has not yet delivered any real-world prototypes. However, we can say with great certainty that they’ll be sticking to a bipedal design, similar to Tesla and Sanctuary.

This one is still a wild card.

5/ 1X’s Eve

Movement: Wheeled base
Torso: Hinged torso joined to articulating hip and knee joints
End effector type: Two-jawed gripper
Current level of autonomy: Teleoperated by humans
Application: Warehouse labor and security
Estimated max payload: Unspecified

Unlike the other robots here, Eve has already seen deployment as a security robot back in 2022. It’s designed to be operated in numbers by a single operator.

Similar to Prosper’s robot, it has opted for a wheeled base instead of legs. Forgoing the complexity of a bipedal design has meant faster development and time-to-market. A singular jointed leg allows it to bend down towards the floor. However, their arm design and lack of specified payload raise questions.

It can place small wooden blocks into a box, but will it be able to handle the heavier loads required for warehouse labor?

Our take: For now, take payload specs with a grain of salt. These numbers are still changing and don’t account for run-time at max load.

There are two distinct approaches: wheeled and legged. All designs seem to rely on teleoperation by a human pilot, all are electrically actuated and sport similar vision systems.

It’s unclear which of these designs will achieve wide-scale adoption first. It really is neck and neck.

MEMES FOR DESSERT

YOUR DAILY MUNCH

Think Pieces

ChatGPT vs Google Bard: which is better?

Opinion: AI isn’t killing creativity — it’s enhancing it.

Startup News

Xtalpi inks $250M partnership deal to find novel drugs with AI.

An oncology assistant for treating cancer patients.

Leading flavor creator branches into food & beverage forecasting.

Research

Ethics and its role in the future of AI development.

A systematic review of AI for clinical decision support (for a specific use case).

Tools

HelpHub: Add custom ChatGPT to your website. AI chat + search for any website or web app. [Sponsored]

Assembly: transcribe and understand audio with AI models.

Luna: create Linkedin content 10x faster with AI.

Lovo.ai: AI voice-generator.

GitMind: free, collaborative mind-mapping software for brainstorming.

TWEET OF THE DAY

A Polynesian security alliance plans to train military drone swarms with AI.

AI ART-SHOW

Until next time 🤖😋🧠