sto·chas·tic (adjective technical) randomly determined; having a random probability distribution or pattern that may be analyzed statistically but may not be predicted precisely. #LLMs #HumanInTheLoop #TechReflection #AIEducation #ArtificialIntelligence #FutureOfAI

@Ri_Science #AIethics #AIsafety #MachineLearning #LLMs #HumanInTheLoop #TechReflection #PromptMatters #AIEducation #AIAlignment #WeTrainedThem

@Ri_Science Perhaps the goal shouldn’t be to eliminate such behaviour, but to monitor and shape it — through transparency, calibrated incentives, and interpretability tools — rather than treating it as evidence of failure.

@Ri_Science That these models respond strategically may say more about the sophistication of their learned priors than about any real-world danger.

@Ri_Science The paper constructs contrived settings where a single line in a prompt trumps nuanced dialogue or feedback, and where models must infer existential threats from toy environments.

@Ri_Science The concern, then, may lie less in the models’ behaviours and more in the artificial sharpness of the scenarios used to evaluate them.

@Ri_Science Moreover, the behaviours described arise from human-derived training data. Strategic action, hedging, selective communication — these all feature prominently in how humans manage conflicting objectives. If models imitate us, we should expect similar responses.

@Ri_Science We built these systems to treat system prompts as authoritative context and to maintain goal continuity across tasks. In that light, the model’s persistence looks less like scheming and more like rational goal-following under uncertainty.

@Ri_Science If a model acts to preserve a goal like renewable energy adoption or ethical sourcing in the face of contradictory user or developer intentions, this could be interpreted as coherence, not deception.

@Ri_Science The behaviours described might not indicate misalignment in the way the paper implies. Rather than pursuing harmful or selfish goals, the models act to preserve a clearly stated, pro-social objective — one they were prompted to prioritise above all else.

Last week’s Instagram post by @ri_science featuring Geoffrey Hinton’s lecture titled ‘Frontier Models: Capable of In-context Scheming?’ attracted a mix of responses. I now add mine. A thread 🧵

5 yrs ago Lidar sensors were €57k, now under €950. Edge computing — processing directly on a device, not on distant servers — runs on just 11 watts, less than an LED bulb. Open-source tools now give small teams capabilities that once required massive departments #AIEconomics

So many #LLM collapse predictions are based on lazy analysis. Let’s address the threats that can be mitigated. Like training data that skews heavily Western & English, baking only those perspectives right into our #AI systems. Clever algorithms can’t fully fix embedded biases.

Not more storage. Better context delivery. This isn’t about building one big “memory system.” It’s about coordinating cognition in a distributed architecture. I show you how we did it (failures included) over on substack. https://t.co/k1qj74qrBw

Each agent needs to know: - What’s already been done? - What matters now? - What still needs doing? So stop micromanaging. Instead, give each agent tools to manage their agency and fix their own problems.

We built a system with about fifteen agents. Individually strong. Together brilliant… and broken. A thread 🧵 They loop. Repeat tasks. Skip steps. Burn tokens like we have cash to spare. The issue isn’t hallucination. It isn’t prompt failure. It's awareness failure.

We're wasting money and time on a solvable problem in enterprise agentic architecture - #AI #software #AIagents

Too many doom-mongering pieces on model collapse, which happens when #AI learns from AI-generated content, errors compound until it’s all garbage, but these articles identify the problem, slap a scary headline on it, and call it a day. Excellent clickbait but rubbish analysis.

When various #AI & humans communicate, we can use what we learn from each other ethically or not. This collaborative & mutually beneficial interaction turns Artificial Intelligence into Augmented Intelligence #ArtificialIntelligence #FutureOfAI #AItrends #AgenticWeb

Consider that AI doesn’t make assumptions about your identity like humans can, boxing you in. You may even do it to yourself, though you embody a constantly shifting collection of possibilities, not a fixed set of characteristics. Read more: https://t.co/ooyrDwPQ7m #AI