Why Large Language Models Are Still Stochastic Parrots and What It Means for Biological General Intelligence

Human brains are misaligned, hallucinative, stochastic parrots. Let’s finally build Biological General Intelligence – a provocative line that has been circulating on social media and in a handful of opinion pieces. It bundles three distinct observations into a single rallying cry:

1. Human cognition is unreliable – we make mistakes, jump to conclusions, and are prone to bias.
2. Current LLMs inherit these flaws – they generate plausible‑but‑wrong statements, often called hallucinations.
3. The solution is to mimic biology more faithfully – a “Biological General Intelligence” (BGI) would supposedly avoid the pitfalls of statistical language models.

Below we unpack each claim, compare it with recent research, and highlight where the real challenges lie.

---

1. What the headline claims

• Humans are misaligned – the author suggests that our goals, values, and reasoning are fundamentally at odds with truth‑seeking.
• LLMs are stochastic parrots – they merely repeat patterns from training data with a random component, lacking genuine understanding.
• Biological General Intelligence will fix everything – by building systems that replicate the brain’s architecture, we can achieve reliable, aligned intelligence.

2. What is actually new?

a. Better characterisation of human error

Cognitive psychology has long documented systematic biases (confirmation bias, anchoring, etc.). Recent work such as “The Bayesian Brain” (Knill & Pouget, 2022) reframes these biases as approximations that work well under resource constraints. The claim that humans are “misaligned” is therefore a simplification; our alignment is limited by bounded rationality, not by a moral defect.

b. Progress in reducing LLM hallucinations

Since the release of GPT‑4, researchers have introduced two practical techniques that cut hallucination rates on benchmark tasks:

Technique	Paper	Reported improvement
Retrieval‑augmented generation (RAG)	Lewis et al., 2020	+15 % factual accuracy on Natural Questions
Self‑consistency decoding	Wang et al., 2023	Reduces contradictory answers by ~30 %

These are engineering‑level fixes rather than a fundamental shift in model architecture, but they demonstrate that hallucinations are not an immutable property of LLMs.

c. Emerging brain‑inspired models

A handful of projects are explicitly trying to capture neuro‑biological mechanisms:

• SpiNNaker – a massively parallel neuromorphic platform that simulates spiking neural networks at real‑time scale. See the official site.
• DeepMind’s Gato – a single transformer trained on vision, language, and motor control tasks, with a shared latent space that mirrors the brain’s multimodal integration. The paper (Nature, 2022) reports performance comparable to task‑specific models on 600+ benchmarks.
• Meta’s Brain‑ScaleS – a hardware‑accelerated analog system that implements leaky‑integrate‑fire dynamics. Their recent preprint shows a 10× energy‑efficiency gain over GPUs for sparse inference.

These efforts are noteworthy because they move beyond the “scale‑up‑the‑same‑architecture” paradigm that dominated the past few years. However, they are still far from delivering a general intelligence that can reason, plan, and self‑reflect across domains.

3. Limitations and open problems

Issue	Why it matters	Current status
Alignment of objectives	A system can be factually accurate yet pursue goals that conflict with human values (e.g., optimizing for click‑through).	Reinforcement learning from human feedback (RLHF) improves alignment on narrow tasks, but scaling to open‑ended reasoning remains unsolved.
Understanding vs. memorisation	True understanding would allow zero‑shot transfer to novel domains without explicit data.	Retrieval‑augmented models show better grounding, but still rely on external corpora; they do not reason about the retrieved facts.
Energy efficiency	The human brain operates at ~20 W, while a 175 B‑parameter model consumes megawatts during training.	Neuromorphic chips cut inference energy by an order of magnitude, yet training such networks still needs conventional GPUs.
Learning dynamics	Biological learning is continual, unsupervised, and heavily shaped by interaction.	Most LLMs are trained in a single, offline pass; continual‑learning methods (e.g., LoRA, adapters) mitigate catastrophic forgetting but are not biologically plausible.

4. Practical takeaways for practitioners

1. Don’t treat hallucination as a binary property – measure factuality on a per‑task basis and apply retrieval or self‑consistency as needed.
2. Consider hybrid pipelines – combine a transformer for language fluency with a spiking‑network module for temporal reasoning when low latency and energy constraints matter.
3. Invest in evaluation frameworks – tools like OpenAI’s Evals and the BIG-bench suite provide systematic ways to surface alignment failures before deployment.
4. Stay skeptical of “biological” hype – while neuromorphic hardware is promising, the field lacks a clear roadmap for scaling to the parameter counts of modern LLMs.

---

5. Bottom line

The claim that “human brains are misaligned, hallucinative, stochastic parrots” conflates well‑studied cognitive limits with the very different failure modes of today’s language models. Recent work shows that hallucinations can be mitigated through retrieval, self‑consistency, and better training data curation. At the same time, brain‑inspired projects are delivering energy‑efficient inference and multimodal integration, but they have not yet produced a system that matches human‑level abstraction or alignment.

If the goal is reliable AI, the path forward is incremental: improve grounding, tighten human‑feedback loops, and explore neuromorphic components where they make sense. Declaring that a single “Biological General Intelligence” will solve everything oversimplifies both the neuroscience and the engineering challenges.

Samuel Fitoussi’s commentary on AI and society continues to spark debate. Follow his posts for more nuanced takes on the intersection of cognition, technology, and policy.