Instagram story viewer> @dailytechwizard> Posts

Daily Tech Wizard

@dailytechwizard

1

followers

0

following

🤖 AI + robotics news in plain English
⚡ Don't get left behind
🧙🏽‍♂️ Grab the Wizard's free forever newsletter ↙️

POSTS STORIES REELS TAGGED

Teams of AI agents cost between 1.8x and 5.1x more than single agents, and they barely beat them.

Evals company Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the Vibe Code Bench. Each model worked solo and as a team, at medium and maximum reasoning effort.

Only one of the four team-versus-solo comparisons showed a statistically significant improvement: GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher. At maximum reasoning, the team setup gave neither Sol nor Opus 5.5 any real advantage. The results suggest the extra cost isn't worth it in most cases, especially when models already run at full compute.

Anthropic saw quality gains shrink as it added more agents in two of its own tests with Opus 5.5. On a knowledge base task, the score went from 0.53 with one agent to 0.70 with ten, then 0.74 with 100. On Lean theorem proving it went from 0.39 to 0.66 with ten agents, and 0.68 with 100.

OpenAI researcher Noam Brown said on the Dwarkesh Podcast that multi-agent systems mainly buy speed, not better quality. Four agents solved tasks twice as fast but also cost twice as much. He said it depends on the task: web research and math parallelize well, but writing a novel doesn't. OpenAI developer Eric Provencher has warned against agent swarms for the same reason, calling the problem a coordination tax.

Would you pay several times more for a team of agents?

Source: link in the first comment

#AIAgents #Claude #GPT Slide 1 of 6. AI agent teams cost more, gain little Vals AI tested solo agents against teams on the Vibe Code Bench. by @dailytechwizard

0

1

8 hours ago

Me checking the new AI label on my favorite song, then hitting replay anyway

The story: YouTube Music now labels AI-made songs

Full story + source in the first comment.

#YouTubeMusic #AIMusic #Streaming by @dailytechwizard

0

1

10 hours ago

POV: you can finally run a giant open model locally, as long as you own two 512 GB desktops

The story: Meta's new open model runs on two Macs

Full story + source in the first comment.

#AI #OpenSource #Meta by @dailytechwizard

0

1

11 hours ago

A walking humanoid built at UCLA uses a smartphone as its torso, eyes and brain.

The robot is called PhoneBot, and the team from the Robotics and Mechanisms Laboratory (RoMeLa) describes it in a preprint on arXiv. Small research humanoids usually need separate sensors, a vision computer and a lot of wiring. PhoneBot borrows those from the handset: the camera for sight, the inertial measurement unit for balance, the processor to run the neural network that drives the joints, and the touchscreen for status and controls.

The body stands 48.3 cm (19 in) tall and weighs 1.8 kg (4 lb), phone and battery included. It has 13 low-cost servo motors, six per leg plus one that swivels the torso so the camera can follow a person without the whole robot pivoting.

PhoneBot learned to walk in simulation through reinforcement learning. On the real robot, the team demonstrated walking, following a person and standing up from lying face down. The researchers tested four Android phones, including a 2017 Honor 9 and three prepaid models bought for about $50 to $80. All ran the core functions, though camera-based mapping (SLAM) didn't run on the Honor 9.

For now it has no arms and was trained only on flat ground. The team estimates the cost at roughly US$400 excluding the phone, against about $3,500 to $13,800 for other small research humanoids listed in the paper, though those platforms offer different capabilities. The hardware and software are open source.

Would you give an old phone a second life as a robot?

Source: link in the first comment

#robotics #humanoid #opensource Slide 1 of 8. A walking robot with a phone for a brain UCLA's PhoneBot uses a smartphone as its torso, eyes and brain. by @dailytechwizard

0

1

13 hours ago

Pentagon wants 60,000 one-way attack drones

Full story in the first comment

#drones #defensetech #Pentagon by @dailytechwizard

1

1

15 hours ago

Webb found a galaxy cluster from 10 billion years ago that already looks modern.

Full story in the first comment

#JWST #Astronomy #DarkMatter

Video: Archive: NASA, public domain

Cover image: sciencedaily.com by @dailytechwizard

0

1

a day ago

Direct an AI voice actor in text

Full story in the first comment

#AI #TextToSpeech #OpenSource by @dailytechwizard

0

1

a day ago

Drone engineers when the camera is too heavy: fine, we'll navigate like a rat and politely touch everything.

The story: This tiny drone navigates by touch

Full story + source in the first comment.

#drones #robotics #engineering by @dailytechwizard

0

1

a day ago

Microsoft says its next in-house AI chip rack makes up to 7x more output.

Full story in the first comment

#Microsoft #AIChips #Maia

Video: Microsoft Azure on YouTube by @dailytechwizard

0

1

a day ago

A new arXiv preprint says training runs for terminal agents should stay matched to the harness that produced them, not get pooled into one shared dataset.

Its method, CoTrace, keeps harness search and model training as separate steps. On the Tmax promotion split, Qwen3.5-9B went from 78 to 88 solved tasks with supervised fine-tuning, and an online reinforcement variant reached 90. On Terminal-Bench 2.1 and SWE-bench Lite, the authors report that transfer to out-of-distribution tasks depends on harness compatibility.

These are the authors' own results in a single preprint.

Source: link in the first comment

#AIagents #LLM #arXiv

Cover figure: Chen, Jixuan et al., arXiv:2610.10426, CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/), https://arxiv.org/abs/2610.10426 Slide 1 of 10. Meme: POV: your agent practiced in one harness, then got deployed in a sibling harness and forgot how to use a terminal. Picture (AI-generated): A confused small robot holds a wrench the wrong way round in a workshop where the tools are arranged unfamiliarly. by @dailytechwizard

0

1

2 days ago

A Starship that flew to space just came home, after 24 days towed through rough seas.

Full story in the first comment

#Starship #SpaceX #Spaceflight

Video: SpaceX on X by @dailytechwizard

0

1

2 days ago

Tesla's Roadster drove upside down. Tesla claims 0-60 mph in under 1 second.

Full story in the first comment

#Tesla #Roadster #EVs

Video: Tesla by @dailytechwizard

0

1

2 days ago