С интересом читаю срачи про GTA 6.
В целом уже понятно, что хардкорные поклонники серии хотели симулятор простого мудака, а получили симулятор нормального сложного человека. Выглядит как стратегическая смена ЦА, и как будто бы с финансовой точки зрения правильная.
На каких же странных вещах держится кукуха у людей. Кажется, мне тоже надо завести себе что-нибудь странное, а то я уже не вывожу. – at Lisbon, Portugal
2
9
233
Ilya Tagunov retweeted
📣 CFP closing soon! Submit a 30-min talk or 15-min lightning talk before Nov 14 for the first Zephyr Meetup in Porto! First-time speakers welcome. Learn more: hubs.la/Q04DpC3K0
Работа всё ещё не даёт соскучиться. В середине ноября наметилась командировка в Индию, летим знакомиться с местными коллегами. А пока я подаюсь на американскую визу, которая нужна для визитов к нашему высокоиу начальству. И надо ещё как-то обычные рабочие задачи не продолбать. – at Lisbon, Portugal
3
8
183
Ладно, бежать, чтобы оставаться на месте, у меня вроде как-то получается. Только надо теперь найти силы бежать вдвое быстрее, чтобы куда-то попасть. – at Lisbon, Portugal
5
87
Интересно, нормы провоза ручной клади в самолётах пришли к нам из каких-то более светлых времён? Я немало летаю по Европе в последние годы, и всегда происходит вот это "наш самолёт полон, пожалуйста, сдайте ручную кладь в багаж". И даже после этого - долгий тетрис с чемоданами. – at Oberding, Deutschland
1
4
116
Выбрался на конференцию в Прагу. Погода приятная, нормальная такая осень. – at Prague, Czech Republic
7
135
Ilya Tagunov retweeted
The loudest voices stoking fears about AI dangers have made tremendous headway in the past two weeks. AI technology has not taken some unexpected, dangerous turn, but the hype around it — propelled by what appears to be a well orchestrated PR campaign — has drummed up considerable fear. I worry that it represents a setback for our field.
I have written frequently that fears of AI are overhyped. AI’s capabilities can be uncannily human-like and unpredictable, and it’s rational to worry when people who are directly involved express concerns. But I see the problems as a sign of the engineering work that ahead, rather than insurmountable barriers or the sky falling. AI technology continues to advance — which is a good thing! — but technical advances, poorly understood by the public, give those who seek to generate hype repeated opportunities to do so.
First, I don’t see any step up in the risk of human extinction from AI compared to a few months ago. The theories about this remain the same fantastical, science fiction scenarios as a few months ago. The biggest change in AI risk is its cybersecurity capabilities — a topic which we should take seriously — but this, too, will not lead to the end of the world.
The most notable recent event leading to increased fear was when an OpenAI team deployed an agent swarm that hacked into Hugging Face. Much of the popular press contained significant hype. For example, some publications reported that a swarm of 1,200 agents carried out the attack. While this was technically accurate, as I write this, I have about 1,300 processes running on my laptop. Yes, the ability to get large swarms of agents to work in parallel on a task is a significant technical advance, And, in computing, many processes run at the same time. So this shouldn’t be seen as some magical capability.
Additionally, OpenAI’s buggy sandboxing and monitoring processes were key to enabling this incident. Fixing these bugs and putting in place improved monitoring would be appropriate fixes, not pausing AI. There are many well known ways to attack software systems. The main advantage of AI agents is that they are relentless. They will tirelessly try many tactics — and have the patience to chain vulnerabilities together — that previously would have taken an infeasible amount of human effort. But in the long term, I believe the advantage will lie with defenders (because they have more information with which to identify bugs, which they can fix), but the cyber-threat landscape has changed significantly. There are still bottlenecks to identifying and exploiting a vulnerability. AI agents still have to try a lot of things to see what works, and taking these actions takes time and might be detected by defenders. This is why, even though it is now easy to obtain versions of leading open weight models that have had their guardrails removed or weakened, so they will not refuse to try to execute cyber attacks, the world has not ended.
I am also concerned about the anthropomorphization of AI in a lot of reporting, where LLMs and agents are unnecessarily treated as if they were people. If I wield a hammer, miss a nail, and accidentally dent the wall, it’s not the fault of the hammer. The problem lies in how I used the hammer. Similarly, if I prompt an agent and it hacks into someone else’s system, the responsibility lies with me, not the agent.
Of course, we want to build systems that are as safe and predictable as possible. (For example, an unsafe hammer would be one whose head randomly flies off under normal use.) Today’s agentic systems are not predictable, but I see no reason why, by applying sound engineering practices, we won’t be able to make them extremely safe to use. One new element in the forecasts of AI-enabled doom is AI companies disclaiming responsibility for their own products. “I didn’t do it; my out-of-control agent did!” There’s a balance to be struck between the responsibility of the tool maker and the tool user, but when something goes wrong, let’s hold the people building and/or using the hammer responsible, rather than the hammer. (By the way, if you’re worried about AI bioweapon risk, David Bellamy has a great post on why this, too, is overhyped. Briefly, the bottleneck in building a bioweapon is not intelligence, but lab work and manufacturing.)
Pausing AI progress will create much more harm than benefit. First, our adversaries will certainly not slow down. Second, engineering requires discovering problems empirically so we can fix them. If we pause AI by a decade, we will also delay finding and implementing safety engineering fixes by about the same duration.
Of course, the incentive to stoke fears — for regulatory capture, to garner attention, or to make one’s technology seem more powerful — remains the same as before. Disclaiming responsibility is a new one. Taking a hard technical look at the actual risks however, I see little factual basis for the degree of fear that’s been stoked up. We still have hard research and engineering work ahead to improve AI safety, but the beneficial applications continue to vastly outweigh the risks, and we should keep building.
[Original text (with links): deeplearning.ai/the-batch/is… ]
Ilya Tagunov retweeted
TLDR (written by AI because I can't be bothered to write short pieces for short attention spans. Original piece written with my own two thumbs):
Today’s frontier models can’t “copy themselves like a virus.” They require terabytes of weights, specialized inference software, tightly configured clusters, and millions of dollars’ worth of heavily monitored GPUs.
Hijacking that infrastructure would create enormous billing, performance, and operational anomalies and displace whatever valuable workload was already running there.
In no way is that not noticed.
Could this become possible as models shrink and compute becomes more widely distributed?
Sure.
But today it’s a foreseeable, mitigable engineering problem, not spontaneous digital reproduction. This propaganda campaign assumes AI advances while infrastructure, monitoring, security, and economic incentives remain frozen in place.
That’s not how technology evolves.
That’s sci-fi with zero understanding of the real world or actual DevOps.
Tell me you have zero DevOps skills without telling me you got zero DevOps skills.
And LLM is not the I Love You Virus. It is not the same as infecting a web server. It is not the same as hacking a wiki. One is a commodity piece of software that is widely distributed and this capability is a non-trivial leap in hardware/software/capabilities and one does not automatically follow the other.
Currently, it is a massive effort to run these models, usually distributed across multiple very expensive chips, running on specialized inference software, specialized harness software, databases, cloud computers, with tooling and hundreds of install packages, etc.
Kimi K3 is 1.6 TB of space at 2 trillion parameters and you need about 1.5 TB of VRAM, which means you need 8xGB300s which will cost about 350-450K a year to purchase or as much or more to rent.
Now let's say Astra or Mythos or later models come in at 10 trillion parameters, so that puts us at roughly 10 to 20 TBs unquanticized. We are guessing her because these folks reveal nothing about architectures anymore, but they are reasonable guesses based on scaling laws. You would need roughly 23 TB of VRAM to run it and roughly several million dollars to run that one instance a year. And to make it worse, it no longer fits on an 8 way cluster, which means you now need special networking configuration and inference software to run it or you need to infect a Nvidia NVL72 rack-scale server system costs between $2.8 million and $6.5 million.
So the "rogue" model copies itself to the most expensive compute in the world. It's safe to say that hardware is not sitting idle and is being watched closely for downtime or anomalies. In fact, it is likely doing one of the following things:
1) Serving customers at a hyperscaler
2) Serving inference for a well endowed company.
In the first case it will be instantly noticed by the actually good automated monitoring systems of the hyperscaler because there is now something using their infrastructure but not being billed properly and costing them millions of dollars in lost revenue and electricity.
In the second example, the model they were serving on that very expensive hardware is now offline (replaced by the rogue model). The model presumably was connected to a valuable piece of internal software that is now glitching and no longer running.
Nobody notices this?
And so we see that this is where sci-fi meets the little problem of the real world where friction, dust, time, resources limitations and the like come to crush your little fantasies.
To be fair, this *may* be possible in coming years. I assign it a non-zero probability. Breakthroughs in new architecture and smaller models and eventually the wider distribution of the computing substrate that commoditizes could change this materially. But for now, it is mitigable and foreseeable and an engineering problem. In no way will these things be copying themselves around like a virus on your computer in their current form.
It also betrays a lack of understanding of the evolution of complex economic and societal and technological systems, we tend to get mitigations as parallel evolutionary developments. Things do NOT evolve in a vacuum.
So when you get viruses, you get anti-virus software.
When you get DDOS attacks, you get CloudFlare.
There is money in solving problems and things do not evolve in isolation.
Believing that things evolve in isolation, or that one thing changes and all other variables stay the same, is one of the many reasoning errors that destroy people's ability to make accurate predictions.
In addition, when you don't have the requisite knowledge to verify your predictions, things that sound sane and rational to you are, in fact, insane and ridiculous.
Порядком надоела бесконечная паника в ленте по поводу самоорганизующихся цивилизаций AI-агентов и вот этого всего. Это модель с триллионами параметров, она инференсится на серверах OpenAI, они могут каждый токен там мониторить, не то что каждый tool call. – at Lisbon, Portugal
1
1
123
Если они сейчас заламывают руки и плачут, что их новые модели такие загадочные и не поддаются контролю, то одно из двух - либо они не хотят этого делать (это дорого, это сложно, чёрный PR тоже PR...), либо они настолько некомпетентны, что их самих в интернет выпускать нельзя. – at Lisbon, Portugal
1
70
Вот над этим наша команда сейчас и работает. – at Lisbon, Portugal
Multimodal AI is pushing the boundaries of what edge devices need to compute.
A new demo showcases the Qwen 2.5 VL real-time vision-language model running on existing silicon enabled by #MIPS ARC NPX #NPU IP.
Powered by MIPS ARC NPX6, the demonstration highlights efficient AI acceleration for vision-language workloads, with the 48K MAC NPU handling prompt processing and token generation while supporting additional vision algorithms in parallel.
See how workload-optimized #AI processors are enabling the next generation of intelligent edge systems : edge-ai-vision.com/2026/07/2…
2
128
Я не могу больше. Наша международная корпорация не даёт разработчикам локальных админских прав - и при этом аутсорсит всё IT, судя по всему, самым дешёвым и некомпетентным индусам на рынке. В итоге банальная установка IDE превращается в квест на несколько недель. – at Lisbon, Portugal
1
5
292









