Nvidia recently posted this video.

@aipost 🏴

🗣 Altman explains current AI memory is still primitive, comparable to early GPT-2 days.

Future systems will remember an entire life, not just facts but subtle preferences. Small habits and unspoken likes will be learned automatically over time. That level of memory may become one of AI’s most powerful capabilities.

@aipost 🏴

Aru Robot, manufactured by the French company Nio Robotics to do industrial inspection and maintenance.

For robotics, it makes perfect sense to look to nature and the animal kingdom for inspiration. Aru resembles a spider. Futuristic and straight out of a sci-fi film: the future is *now* here!

@aipost 🏴

📢 Unemployment continues to rise sharply; no turnaround in sight!

"As AI-powered automation replaces jobs at a fraction of the cost of human labor, understanding the value of your own skills and whether tools like ChatGPT can outperform them has never been more critical."

- For Gen Z, the entry-level career ladder is getting steeper by the month - and there’s no sign of it letting up. Unemployment among recent grads has climbed to 5.8% (the highest since 2013, excluding the pandemic) as companies rethink hiring amid AI-driven productivity gains.

-"Goldman Sachs executive has a blunt message for young professionals trying to get ahead: Know what you bring to the table."

Source.

@aipost 🏴

You can now adjust specific characteristics in ChatGPT, like warmth, enthusiasm, and emoji use.

@aipost 🏴

📱 FunctionGemma, Google puts an AI agent directly on your phone

Google has quietly released FunctionGemma, a compact AI model that runs entirely on-device and can turn voice commands into real system actions. No internet, no cloud, no data leaving your phone. Release date: December 18, 2025.

What it is
• 270M-parameter model optimized for mobile
• Fully offline and private by design
• Converts natural language into executable phone actions
• No servers, no latency from network calls

How it works
• You speak a command
• The model parses intent locally
• Translates it into structured system functions
• The phone executes instantly

Example: “Add John to contacts, number 555-1234” → contact saved, zero cloud involvement.

Why it’s different
• Traditional assistants rely on cloud round-trips
• FunctionGemma runs at ~0.3s latency on-device
• Privacy by default, nothing leaves the phone
• Works even with no signal

Performance snapshot
• 270M parameters (≈6,600× smaller than GPT-4)
• 126 tokens/sec
• ~85% accuracy after fine-tuning (up from ~58%)
• ~550 MB RAM usage
• ~288 MB file size
• ~0.75% battery drain for ~25 interactions

What it can do
• Set alarms and reminders
• Create calendar events
• Control smart-home devices
• Navigate locally
• Log simple activities

All processed locally. All instant.

Architecture insight
Google frames this as a “traffic controller” model:
• Simple, sensitive tasks → handled on-device
• Complex tasks → optionally routed to cloud AI

Limitations
• No multi-step task chaining yet
• Struggles with indirect phrasing
• 15% error rate without tuning
• Best results require app-specific fine-tuning

This isn’t about one assistant. It’s a shift in philosophy: small, specialized models doing real work locally can outperform larger general models, faster, cheaper, and privately.

AI isn’t just moving faster, It’s moving closer to you.

https://huggingface.co/google/functiongemma-270m-it

@aipost 🏴

Japanese researchers created CirculaFloor, robotic tiles that let you walk endlessly in VR without moving an inch.

@aipost 🏴

Yoshua Bengio says AI does not create enough new jobs to balance the ones it replaces

A handful of engineers earn huge salaries, while a vast number of workers face displacement as models master cognitive tasks. "If we automate most of the cognitive work, what's gonna be left?"

@aipost 🏴

🔥 Google’s quantum processor could perform 1,000 times better with Princeton’s new qubit

Princeton researchers built a superconducting qubit with much longer coherence than today’s standard designs, on the order of milliseconds instead of microseconds.

The improvement comes from using tantalum on high purity silicon, which dramatically reduces energy loss. Longer coherence means fewer errors and far more useful computation.

In principle, dropping this qubit into existing quantum processors could yield ~1,000x effective performance gains, with benefits compounding as systems scale. This is a materials and fabrication breakthrough, not a software trick, and it directly targets one of the biggest bottlenecks in scalable quantum computing.

Source.

@aipost 🏴

OpenAI is now aiming to raise as much as $100 billion at a $830 billion valuation (up from the previously reported $750 billion valuation).

OpenAI's valuation jumped to $500 billion in October 2025 (up from $300 billion earlier this year).

Wild.

@aipost 🏴

San Francisco power outage halts Waymo self-driving cars

@aipost 🏴

Your blue-collar job isn't secure either. Take a close look at how delicately kyber labs' robots perform manual labor. That's nuts.

@aipost 🏴

The AI boom is accelerating:

US tech companies are committing to spend a combined $569 billion on data center leases over the next several years.

A lease is a long-term rental agreement, meaning these companies commit to multi-year payments for data centers, offices, or warehouses, without owning them upfront. This represents a +$197 billion, or a +53% increase from Q2 2025. Oracle, $ORCL, alone added +$148 billion in new lease commitments in Q3, bringing its total data center lease obligations to $248 billion. Some of these leases can last up to 19 years, locking Oracle into massive fixed costs regardless of future AI demand.

Tech companies are making massive bets on AI.

@aipost 🏴

📈 OpenAI’s Compute margins surge again

OpenAI’s compute margin on paying users reportedly hit ~70% in October (up from ~52% at end-2024 and ~35% in Jan 2024), driven by cheaper rented compute, inference efficiency tweaks, and a higher-priced subnoscription tier.

- OpenAI’s compute margin for paying users jumped to 70% in October

- Company had declared a ‘Code Red’ to focus on server costs after release of new model from DeepSeek

- For total computing costs, Anthropic is expected to be more efficient

@aipost 🏴

🇨🇳 China is deploying Robots everywhere.

Here Unitree's humanoid robots + quadruped robot dogs patrol and perform duty in a team.

Robots handle patrols, officers focus on judgment and de-escalation.

@aipost 🏴

📈 US is the absolute king of total available AI compute at 74.4%.

China at 14.1% and
the European Union is at 4.8%

So US has 5.3x the capacity of China, and 15.5x of EU.

@aipost 🏴

Elon Musk on X.

@aipost 🏴

Your Year with ChatGPT!

OpenAI is rolling out You Year on ChatGPT to everyone in the US, UK, Canada, New Zealand, and Australia who have reference saved memory and reference chat history turned on.

Just make sure your app is updated.

@aipost 🏴

🔥Paralyzed Patient in China Controls Robots Using a Brain Implant

For the first time, a person with a severe spinal cord injury has achieved stable, real-world control of multiple assistive robots using an implanted brain–computer interface (BCI). The breakthrough was reported in China and involves a patient who has been completely paralyzed from the neck down for three years.

Using the implanted BCI, the patient can remotely control different robotic systems in everyday situations. These include ordering food online and sending a robot dog to meet the delivery courier, performing paid digital work, and moving around the neighborhood via robotic assistance.

Unlike earlier laboratory demonstrations, the system works reliably in daily life rather than controlled experimental settings. The achievement highlights a major step forward for brain–computer interfaces, suggesting that BCIs may soon enable people with severe paralysis to regain practical independence and participate more fully in work and social activities.

@aipost 🏴

A humanoid robot performing tactical rolls and recovery moves

@aipost 🏴

Physical intelligence: Fully autonomous, not teleoperated.

@aipost 🏴