Digest Topics Claims Predictions Reliability Researchers

Digest Topics Claims Predictions Reliability Researchers

HypeDelta - AI Research Intelligence

Topicsscaling

scaling

62% bullish

135 claims over the last 90 days

Total Claims

135

Lab Researchers

16

Critics

64

Other

55

Avg. Sentiment

Bullish

Lab Researcher Claims

What researchers at major AI labs are saying

fact

Bearish

Base LLM scaling reached a capability plateau between 2022-2024

Francois Chollet

8/9/2026

opinion

View all claims for this topic

Bullish

After o3's test-time compute demonstration in late 2024, the new models showed genuine fluid intelligence and the LLM line of research could achieve unbounded capability scaling with no wall

Francois Chollet

8/9/2026

prediction

Neutral

Future AI in 15 years will not be based on the LLM stack but will necessarily move closer to symbolic learning as its optimal final form

Francois Chollet

8/9/2026

opinion

Neutral

Current AI techniques are 4-6 orders of magnitude away from optimality in terms of data efficiency and test-time compute efficiency

Francois Chollet

8/9/2026

fact

Bullish

RL improves reliability more than coverage

8/8/2026

fact

Neutral

At low total compute budgets, pretraining dominates because weaker policies benefit less from RL

8/8/2026

fact

Neutral

The optimal compute split between pretraining and RL changes with scale, shifting from roughly 20% to 28% RL as budget grows

8/8/2026

fact

Bullish

Stronger pretraining makes RL itself scale better, with pretraining loss determining the performance that can be reached at a given RL budget

8/8/2026

opinion

Bearish

Deep learning may be hitting a wall in its progress

8/2/2026

Critic Claims

What critics and skeptics are saying

opinion

Bullish

The book 'Build a Large Language Model (From Scratch)' is a best seller and receives a five-star rating.

9/1/2026

fact

Neutral

Pre-training with limited data exhibits different scaling behavior than web-scale language modeling, where increasing parameters beyond an optimal point causes overfitting rather than improved performance

Machine Learning

8/29/2026

fact

Neutral

Optimal model size depends strongly on both the data budget and the downstream target task

Machine Learning

8/29/2026

critique

Bearish

Standard Transformers scale down poorly to limited data settings because embeddings consume too large a fraction of parameters and per-token computation is coupled with representational capacity

Machine Learning

8/29/2026

fact

Bullish

Recursive Transformers with shared blocks and factorized embeddings outperform standard Transformers on limited data budgets of 10M and 100M words

Machine Learning

8/29/2026

fact

Bearish

AI industry has current revenues in tens of billions or low hundreds of billions against Capex in the trillions

8/28/2026

fact

Bearish

More ambiguous next-token distributions are harder for LLMs to learn accurately

Neural and Evolutionary Computing

8/28/2026

fact

Bearish

More ambiguous distributions require more capacity, larger embeddings, more training steps, and amplify sampling noise

Neural and Evolutionary Computing

8/28/2026

fact

Bullish

Scaling post-training alone was sufficient to achieve GLM-5.3's improvements over GLM-5.2 using the same base model

8/28/2026

prediction

Bullish

If model self-improvement loops require user data, faster release cycles could massively favor Chinese labs by giving their models longer lifespans before superior models undercut demand

8/28/2026

Other Claims

Independent, journalist, and unclassified remainder

fact

Bullish

Every year since 2022, one more component of the pipeline that produces machine intelligence has flipped from human-made to model-made

8/28/2026

fact

Bullish

The reward signal was the first thing to go synthetic in 2022 with InstructGPT and Constitutional AI

8/28/2026

opinion

Neutral

Parameter count is only meaningful alongside data quantity, compute allocation, and deployment conditions

8/28/2026

fact

Bullish

Advanced skills like finding software vulnerabilities require carrying long causal chains of 20+ inference steps and do not live in total parameter count once a knowledge-holding threshold is reached

8/28/2026

opinion

Bullish

As agent capability improves, much of the difficulty in scaling post-training moves from the model to the environment

8/28/2026

fact

Bullish

Ornith-1.5 achieves end-to-end self-improvement where the model proposes tasks, generates scaffolds, and produces RL rollouts to create new training experiences

8/28/2026

fact

Bullish

Qwen 3.8 Max is a 2.4T parameter model that would have been the top open model except for Kimi K3

8/4/2026

fact

Bullish

Qwen models can perform autonomous coding for 10+ days unattended and build complete systems from scratch

8/4/2026

fact

Bullish

AI models can autonomously conduct research and improve upon published papers, achieving +2.71 point improvements

8/4/2026

prediction

Bullish

Qwen will open-weight both the 3.8 Max model and related models

8/4/2026