What is happening in the industry right now, and why everything is speeding up again
Artificial intelligence has entered another phase of rapid acceleration. The industry giants are revealing hidden projects, research labs are breaking the old limits of models, and start-ups are showing results nobody expected of them. Below is a detailed digest of the latest events to help you see the bigger picture.
1️⃣ OpenAI declares an internal “code red” and speeds up work on its hidden model, Garlic
After the rapid success of Gemini 3, OpenAI declared an internal “code red”: a signal that the company sees a real threat to its lead.
The main news from the leak was the existence of a closed model called Garlic, an internal project that had not been disclosed publicly before. It includes fixes to the pre-training process, which are meant to let the model:
- first build broad, generalised connections,
- then gradually add detail,
- avoid the “data overfeeding” effect,
- work more efficiently with fewer parameters.
According to internal tests, Garlic:
- outperforms Gemini 3 and Claude 4.5 in reasoning and complex programming;
- shows denser “knowledge packing”;
- achieves strong results without a sharp increase in parameters.
2️⃣ Alibaba LiveAvatar: a system that creates talking avatars in real time
Alibaba has unveiled LiveAvatar, a video model with ~14B parameters that can:
- generate an avatar in real time: 20 frames per second on five H800 server GPUs,
- run more than 10,000 seconds (almost three hours) of continuous streaming without degradation,
- sync facial expressions and lip movements with speech,
- work together with Omni language models as a fully fledged live agent.
The main technical breakthrough is stability over ultra-long videos, something other generative systems could not manage.
Alibaba solved the problem with three techniques:
- Rolling RoPE: dynamic updating of the reference frame, which preserves identity.
- Adaptive Attention Sink: replacing the reference frame with a generated one to correct drift.
- History Corrupt: adding noise so the model learns to recover from errors on the fly.
The result is avatars that can hold long conversations without “melting faces” or visual artefacts. That makes LiveAvatar a serious step towards the digital assistants of the future.
3️⃣ Anthropic study: AI saves creatives time, but colleagues still look askance
Anthropic launched a system of automated interviews and ran a study of 1,250 people, including scientists and people in creative professions.
The main findings:
- 97% of creatives say AI saves them time.
- 2/3 believe it improves the quality of their work.
- 70% said they have to manage how colleagues perceive their work with AI: in creative circles it still carries a stigma.
- Respondents describe 65% of their AI use as assistance, while across conversations with Claude as a whole, automation accounts for about half of tasks (49%).
Examples from the answers:
- a copywriter raised their output from 2,000 to 5,000+ words a day;
- a photographer cut editing time from 12 weeks to 3.
Scientists are calmer about AI: 91% want it as a research partner, but they still insist on double-checking every line of reasoning.
4️⃣ Smart street lights as mini data centres: a new kind of urban infrastructure
The British company Conflow Power has proposed turning street lights into distributed computing nodes.
The iLamp street lights are fitted with:
- solar panels,
- batteries,
- traffic and air sensors,
- 5G modules,
- and built-in Nvidia Jetson units.
The idea is to create a network of micro data centres across the city that can:
- handle local AI tasks with minimal latency,
- reduce the load on conventional data centres,
- earn money for the owners of the infrastructure.
If the project takes off, cities will quite literally become “smart networks of computers”.
5️⃣ Microsoft VibeVoice: a voice that responds in under 0.3 seconds
Microsoft has unveiled a compact speech synthesis model that:
- starts speaking ~300 ms after receiving the text,
- runs in parallel with an LLM,
- keeps speech quality on a par with large TTS models,
- holds its quality over long monologues.
The architecture includes:
- a purely acoustic tokeniser,
- a small diffusion head,
- a 0.5B-parameter language model (Qwen2.5-0.5B).
The effect is an almost natural conversation, without the familiar pause before a voice assistant answers.
6️⃣ Smart glasses that “reveal” the life stories of passers-by
A video has gone viral showing glasses that:
- recognise a person’s face,
- find information about them in open sources,
- and produce a mini-dossier within a few seconds.
Passers-by react with anything from laughter to fright: the technology shatters the familiar idea of anonymity in real life.
This is not science fiction but a combination of three technologies that already exist:
- facial recognition;
- searching open data;
- and compiling a short profile with an LLM.
It is one of the most hotly debated flashpoints between AI and privacy.
7️⃣ Humanoids: from factories to combat sparring
⚡ EngineAI T800
A demonstration sparring session ended with the robot striking the company’s chief executive and knocking him to the floor, even through thick protective gear. The video set off a heated debate about the safety of autonomous systems.
⚙️ Midea MIRO U
A six-armed industrial robot on a wheeled platform:
- can carry out several operations at once,
- rotates 360° on the spot,
- can change tools on the move.
Midea planned to launch a pilot at its washing machine plant in Wuxi before the end of 2025.
8️⃣ AI Safety Index: nobody is making AI truly safe
An independent assessment of the industry painted an unpleasant picture:
- no company scored higher than C+,
- the best grades were a C+ for Anthropic and OpenAI and a C for Google DeepMind: average, and far from an acceptable standard,
- in existential safety everyone failed,
- Chinese companies got grades of around D/D–.
The experts’ conclusion:
the industry is moving towards AGI faster than it understands how to control it.
9️⃣ Google Titans and MIRAS: models that can remember millions of tokens
Google has presented two major lines of research.
🔵 Titans
A long-term memory mechanism that:
- combines windowed attention with an updatable memory,
- chooses what to remember based on how surprising it is,
- lets models with <1B parameters beat the giants on long-text tasks.
🔵 MIRAS
A single theoretical framework for all sequence architectures.
On its basis Google built attention-free models (Moneta, Yaad, Memora), which outperform Mamba 2 on a number of tasks.
🔟 Google versus the media: AI has started rewriting article headlines
Users noticed that the system was replacing headlines in Google Discover so aggressively that it changed the meaning of the articles.
The problem:
- the “AI-generated” label is visible only after a click,
- the publications’ logos stay in place,
- readers blame the journalists, not the AI.
Media outlets have already protested: it undermines trust in the news and distorts the information space.
1️⃣1️⃣ Lux: an agent that operates a computer better than the giants’ agents do
The start-up OpenAGI Foundation has shown an agent called Lux, which can:
- read interfaces from screenshots,
- click, type and work with files,
- operate a full desktop, not just a browser.
On the Online-Mind2Web benchmark, according to the company itself, Lux achieved an 83.6% success rate, higher than the solutions from the market leaders.
The trick is learning from actions: best experience, then new data, then a better strategy.
1️⃣2️⃣ The X4 mini reader: an E-ink display that attaches to your phone
A Chinese gadget that turns a smartphone into a compact e-reader:
- a thin 5.9 mm E-ink panel,
- MagSafe magnets,
- up to two weeks of use on a charge,
- books loaded over Wi-Fi and microSD.
In effect it is an e-reader “sidecar” for your phone, with no extra apps or notifications.
1️⃣3️⃣ A ring for traders: vibrations when the market moves
An unusual accessory integrated with TradingView:
- the ring vibrates when assets move sharply,
- changes colour depending on market dynamics,
- lets you “feel the market” without a phone.
There is no information yet on the price or when sales start, but the concept has drawn a lot of interest.
1️⃣4️⃣ Lyria Camera: an app that turns reality into music
Google has shown an experiment:
- the camera reads the scene,
- Gemini describes it,
- the Lyria music model creates a live soundtrack,
- the music adapts in real time to what the camera (or the computer screen) sees.
The developers call it “a new kind of musical interaction with the world”.
1️⃣5️⃣ Poetiq: an ARC-AGI-2 record without a giant model of its own
The start-up Poetiq set a historic record on the ARC-AGI-2 benchmark: 54% accuracy, passing the 50% mark for the first time.
The key point is that it did this without a large model of its own.
In essence, Poetiq built a “meta-system” that:
- combines several models,
- iteratively improves the answer on top of existing models: it generates a solution, checks it and corrects it until a self-verification module decides it is good enough,
- picks reasoning strategies for each task.
It is a step towards systems where intelligence comes not from the size of a model but from smart coordination of many models.
1️⃣6️⃣ OpenAI raises its bet on the enterprise market
What is happening with ChatGPT Enterprise, according to the company itself:
- weekly messages are up roughly 8-fold,
- use of Projects and custom GPTs has grown 19-fold,
- reasoning token consumption per organisation has risen 320-fold, as tasks get more complex,
- companies are building thousands of internal assistants.
Because of its enormous infrastructure costs, the enterprise segment is becoming critically important for OpenAI.
What to look for as a user
Tell a demo apart from a working product. For a real use case, what matters is access to the right data, limits on permissions, a log of actions and the ability to stop the agent. The more autonomy a system gets, the more important it is to define the boundaries of human responsibility in advance.
Useful articles and discussion here (chat in Russian)
