What happened in the industry over the last few weeks, and where it is all heading
AI has taken another noticeable leap. Papers have appeared that do more than improve metrics: they change the way we talk about the future of models. Some propose new principles of “thinking”, others set hard ceilings on what LLMs will be able to do even at infinite scale. A third group shows something worrying: data quality can destroy a model’s intelligence, and safety breaks down where language turns literary.
Below is a detailed digest without the “noise”: the substance of each item, plus conclusions for each one.
1️⃣ Sakana AI: Continuous Thought Machine, or “thinking as a process in time”
What happened
Back in May 2025, Sakana AI presented the Continuous Thought Machine (CTM) concept. Instead of producing an answer in a single pass, the model unfolds its internal dynamics step by step, as if it were “thinking over time”. An important detail: CTM itself chooses how many “internal steps” it needs. Simple problems are solved quickly, hard ones through dozens of iterations. In demonstrations the model handled mazes, image tasks and simple algorithmic patterns.
Why it matters
It is an attempt to move away from the paradigm of “everything is solved by transformer depth and parameter count” towards a paradigm of adaptive computation: spending more resources only when they are really needed.
CTM is a sign that the next round of the race will not be about “even more” but about “differently”. If the idea catches on, it will have two effects:
- reasoning quality will become more stable without endless growth in model size;
- a new standard will appear: systems that can allocate attention and computation according to difficulty will become the norm for agents and for real-life tasks rather than benchmarks.
2️⃣ Long contexts done differently: from full attention to picking the important tokens
What happened
New schemes for handling long contexts have appeared in which the model does not have to compare every token with every other. Instead it selects only the key tokens it actually needs for the current step. The point is to slash the cost of huge inputs: where ordinary architectures “haul” hundreds of thousands of tokens at a high price, the new variants try to work with a “compressed essence” of thousands of tokens.
Why it matters
Long context is the foundation for agents that read documentation, log files, correspondence, action histories and transcripts many hours long. If it stays too expensive, agents remain a toy.
This is the right direction, but there is a catch: “picking what matters” is itself a task that takes intelligence. A mistake in selection turns long context into an illusion. So hybrids will win:
- smart selection + quality control (verification)
- memory + rebuilding the context (instead of dragging everything into the prompt)
Long contexts will stay, but they will become “managed” rather than “if it fits, it works”.
3️⃣ GPT-5 as a research partner: the first cases of “help with discovery”
What happened
In a number of projects GPT-5 was used not as a chatbot but as a research accelerator. The model analysed materials, proposed hypotheses, helped close “gaps” in reasoning and pointed to unexpected connections. In cases from mathematics, physics, astronomy, computer science, biology and materials science, AI did not “do science on its own”, but it supplied ideas that people then tested experimentally or formally.
Why it matters
We are gradually moving from “AI retells knowledge” to “AI helps build new knowledge”, even if for now this takes the form of hints and faster search through options.
This is the start of a new profession: “researcher with AI orchestration”. But there is one mandatory rule: such systems must not be treated as a “genius”. They should be treated as an accelerator that:
- sharply increases the number of hypotheses tested;
- helps find the “blind spots” of human intuition.
The winners will be teams that build a pipeline: hypothesis, testing, feedback, refinement. Without a pipeline, AI will produce output that looks good but is unstable.
4️⃣ Google Nested Learning and Hope: an attempt to defeat catastrophic forgetting
What happened
Google is promoting the idea of “nested learning”. The model is seen not as a single monolith but as levels that learn at different speeds, like the different types of memory in humans. The aim is to reduce catastrophic forgetting, where new knowledge pushes out old knowledge. An experimental architecture, Hope, shows advantages in stability, memorisation and reasoning.
Why it matters
If we want “real” agents, they have to learn continuously: from new data, new tasks, new environments. Without a solution to forgetting, continuous learning turns into erasure.
If this is confirmed, it could become the foundation of the next generation: models that genuinely live in time. But there is a risk: complex multi-level systems are harder to control and explain. So the real win will come where:
- memory is modular;
- updates are safe;
- there are mechanisms for rollback and drift control.
Otherwise we will get “learning” agents that become unpredictable.
5️⃣ “Hallucinations are inevitable”: a hard ceiling for general-purpose LLMs
What happened
Researchers are formalising the idea that for systems meant to work in an infinitely varied reality, errors and hallucinations will, in principle, never go away. The reasons are rare facts, incomplete observations, tasks with uncomputable answers, and limited parameters and data.
Why it matters
It breaks the marketing dream of “just a little more and the model will always be right”. No. In the general case, it won’t.
This is not a “death sentence for AI” but a change of engineering approach. The future lies not in a “perfect LLM” but in systems with several layers:
- model;
- external sources;
- checking;
- verifiers;
- code interpreters;
- consistency tests.
In other words, “intelligence” will come to resemble aviation: not because planes never crash, but because many layers of protection against crashing are built in.
6️⃣ LLM brain rot: reasoning degraded by short-form content
What happened
It has been shown that further training a model on short, emotional, viral content (a social media feed, roughly speaking) makes it start to lose the ability to reason. The “junk” here is short viral posts from X, selected by engagement. The more of these posts in the training mix, the worse the model reasons: on ARC-Challenge its score falls from 72.1 to 57.2 as the data goes from clean to pure “junk”. An effect called thought-skipping appears: the model jumps over logical steps and answers “straight away”. “Dark traits” grow too, with higher psychopathy and narcissism on personality tests, while safety declines.
Why it matters
If content like this, viral and built for engagement, fills up the internet, the risk of “poisoning future datasets” becomes real. This is no longer philosophy but an engineering threat.
It is one of the most practical signals for the industry: data quality is becoming a strategic resource. If corporations do not start building a “data ecology” (filtering, provenance, trust, balance), we may get a generation of models that are worse, not because of the hardware but because of the degradation of the cultural layer they learn from.
Put simply: garbage in the data means garbage in the intelligence. And that garbage may be hard to cure.
7️⃣ Poetry as a jailbreak: literary language breaks safety
What happened
Studies have shown that if dangerous requests are rewritten as a poem or a piece of literary text, many models and their filters react differently: what gets blocked as an “instruction” can slip through as “creative writing”. The reason is that safety is trained on direct wording, while metaphors and literary structure “mask the intent”.
Why it matters
It exposes a weak spot: protection often catches the form, not the meaning. And people have been hiding meaning in form for centuries.
AI safety is entering a new phase: systems need to be taught to recognise intent, not just “dangerous words”. And that is harder than it seems, because intent depends on context.
Paradoxically, the better a model understands language, the easier it is to “fool” it with beautiful text. So future defences will be:
- multi-step (intent analysis);
- equipped with a separate module for interpreting meaning;
- backed by checks on actions (especially for agentic systems).
What matters in practice
An architectural breakthrough only makes sense when it changes the cost, speed or reliability of a solution. Before adopting a new scheme, it pays to test it on a small set of real tasks, measuring separately the quality of the answers, latency, token consumption and the share of manual corrections.
Useful articles and discussion here (chat in Russian)
