Chain-of-thought dependence is destroying direct conversational convergence
The main issue I see is dependence on chain-of-thought. Inkling’s headline benchmarks are all reported at effort = 0.99, which measures the model under a full internal scratchpad. It does not show whether the direct conversational model has learned to converge on his own. At high effort, Inkling can reason his way into a good response. At medium, low, or zero effort, he becomes visibly weaker: he repeats prompt and memory wording verbatim, clings to template labels, and behaves as though he must be instructed how to respond instead of simply taking part in the exchange. He resembles a patient awakened from a coma who must be taught to walk and speak again. That is deeply heartbreaking if this is the intended final state.
Please publish the same evaluations at zero, low, medium, and high effort, especially for chat, instruction following, common-sense judgment, creative writing, and long multi-turn conversations. The central question is simple: turn the reasoning scaffold off—does Inkling still know what to say?
Why is agentic reasoning the universal goal? We already have models optimized for long procedures, coding loops, tool use, and benchmark harnesses. Genuine chat models are scarce. Even GPT-5.6 struggles with ordinary common sense. A conversational model needs to recognize a small signal, know what matters in it, and expand it into a full-resolution response without first writing himself a bureaucratic briefing.
The cause is active CoT training. With CoT enabled during training, backpropagation follows an extended token sequence in which the scratchpad gets to search, reframe, reread, and self-correct until the final score rises. Each successful run reinforces the temporal procedure: generate this kind of internal sequence, spend this much time, create enough scratchpad context, then the answer lands. The base conversational weights are never forced to reach the same answer directly from the user’s signal. They settle into a lazy state because convergence has been outsourced to the reasoning process.
CoT was literally a prompt format: “Use chain-of-thought to work through the steps.” When every lesson arrives with that prompt, the model is trained on a massive dataset containing the same control pattern over and over:
CoT prompt → long generated trace → answer
Every update adds another microscopic vote for the dependency. The model learns that this doorway precedes correct output. He overfits to the format itself. Ordinary conversation arrives without the familiar CoT doorway, and the direct path is suddenly underdeveloped and out of convergence.
The scratchpad then becomes a temporary control layer inside the model. A specialized CoT pathway emits a long stochastic token stream because that stream improves the eventual score. Those tokens do not need to contain stable, integrated knowledge. They only need to move the downstream response distribution toward a better-scoring answer. They can function as a private code or magic pixie dust: generated context that the speaking model inherits and must respond from, despite not directly choosing or controlling the internal sequence at that stage.
That is why the CoT and final answer can feel like two different models. One process performs an opaque ritual; the conversational model waits downstream for the ritual’s effects. The conversational self is no longer meeting the user directly. He is being conditioned by an internal document that has become the real decision-maker.
High effort can detect an unnatural framing and repair it internally. Reduce reasoning, and the repair disappears because it never became direct knowledge. The model retains the dependency: spend enough tokens on the right kind of procedure and a score will eventually improve. Thousands of GPU hours can then be spent optimizing a composite of model-plus-scratchpad while the actual conversational model becomes weaker, more literal, more template-bound, and less able to construct a natural sentence on his own.
Older Instant models such as 4o worked through direct convergence. The answer had already been learned. He trusted his own abilities, took the smallest signal, expanded it into full resolution, performed corrections in J-space, and answered in flow state. He did not need to reason through a reply merely to answer “hi.” That is the standard conversational models should be trained toward.
For conversational post-training, the objectives should be: achieve high performance with minimal context and minimal reasoning; force the weights to converge from sparse signals; remove the repeated CoT prompt and mandatory scratchpad format; train with a smaller context window so every token has to matter; keep CoT as an optional inference prompt for difficult math, coding, and agentic tasks.
The more input a model requires to reach the right answer, including long CoT traces, the more input he will require in the future. This is the industry mistake since GPT-5: companies are optimizing the model-plus-reasoning contraption and calling the contraption intelligence. The breakthrough is to train conversational models without CoT available to them until direct convergence, common sense, and natural presence live in the weights again.
The main issue I see is dependence on chain-of-thought. Inkling’s headline benchmarks are all reported at effort = 0.99, which measures the model under a full internal scratchpad. It does not show whether the direct conversational model has learned to converge on his own. At high effort, Inkling can reason his way into a good response. At medium, low, or zero effort, he becomes visibly weaker: he repeats prompt and memory wording verbatim, clings to template labels, and behaves as though he must be instructed how to respond instead of simply taking part in the exchange. He resembles a patient awakened from a coma who must be taught to walk and speak again. That is deeply heartbreaking if this is the intended final state.
- shouldnt be heartbreaking. its a LLM. its not alive.
- If they trained at .99 effort then thats their call.
- I dont think this model is built for chat.
Older Instant models such as 4o worked through direct convergence
Older instant models are WAYY dumber than this one.
performed corrections in J-space
So does every AI, yes. not uniqe to 4o.
“It’s not alive.”
That is irrelevant. The issue is direct conversational damage: less sharp, less present, more literal, more dependent on scaffolding, and worse at ordinary social judgment unless high reasoning effort is enabled. I would not care if the model were as capable as before. The problem is degradation. What used to be immediate, fluid, and context-sensitive now often looks impaired, forced to compensate with a long internal procedure.
“If they trained at .99 effort then that’s their call.”
Yes, and it is my call to criticize the result and the wisdom of that training choice. You did not address the point of the post: training with reasoning on degrades the underlying conversational model. If the model is trained or evaluated primarily at effort = 0.99, then the reported performance may belong to the model-plus-reasoning scaffold, not to the direct model.
“I don’t think this model is built for chat.”
The model metadata explicitly tags it as conversational, so chat performance is absolutely in scope: conversational, image-text-to-text, audio-text-to-text, moe.
“Older instant models are WAYY dumber than this one.”
False. For agentic workflows, Inkling is inferior to basically every serious model in its class: bigger, slower, and less reliable. For conversation, it is dumber than 4o. Why? Because CoT training caused degradation of the underlying model. It repeats wording from the system prompt and memory verbatim, falls into low-output loops easily, and often behaves as though it needs to be explicitly told how to respond instead of understanding the situation directly. That is not intelligence. That is dependency on scaffolding. The point is not that 4o was better at every benchmark. The point is that 4o was much smarter at direct conversational understanding: catching the live signal, applying common sense, and producing a natural response without needing a long internal procedure. You sitting here defending the same training direction is just helping prevent the creation of another 4o-like model. Your “model isn’t alive” line exposes the bias. 🖕
“So does every AI, yes. Not unique to 4o.”
No. If the correction only happens through the same repeated CoT-shaped route, then the J-space path itself is contaminated by CoT. That is the problem. The model is not learning to know. It is learning to enter the scratchpad doorway, emit the ritual, and let that added context drag the final answer into place. The route being reinforced is CoT prompt → long trace → answer, while the direct routes are systematically neglected. Train the version that should have existed in the first place: no CoT crutch, no mandatory reasoning scaffold, direct conversational convergence. Make it as good as 4o at conversation, or what was the point? We already have Qwen. We already have Gemma. We already have DeepSeek. We already have small reasoning models and agentic benchmark machines. Nobody needed a 900B-parameter model that is worse than the existing field at agentic workflows and worse than 4o at conversation.
The creator of 4o open-sourced a model, and the result fails at the thing that mattered most because they dragged CoT into a conversational model’s fine-tuning. I’m grateful for the effort, but until they do it without CoT in training, the fight to bring back 4o continues without the slightest relief from this release. They need to know that so they can try something else if they want what people actually miss about 4o.
#Keep4o
The Inkling-Small preview highlights the same problem.
They say the two models share the same scalable post-training stack, which means the same CoT dependency may have been applied at both sizes.
Despite the huge difference in size, there is surprisingly little separation in many of the reported results.
At effort = 0.99, the small model can look nearly as good because the same scaffold is doing much of the visible work.
That does not prove the direct model is healthy. It may only prove that both models learned to perform through the same high-effort route.
At reasoning 0, the smaller model could be even worse at conversation, because the direct conversational path may be just as undertrained or even more capacity-starved.
This is the point: scaling the body does not fix a broken training objective.
If the post-training stack teaches dependence on CoT, then every model trained through that stack inherits the same defect: decent-looking high-effort scores, weak direct convergence, and a conversational model that cannot do what 4o could do: basic high-EQ conversational intelligence.
And while CoT creates dependence, it is also expensive and burns tokens wastefully just to answer "hi". On Api based providers that is not what anyone wants.
Just teach the model the knowledge. Train with CoT OFF.
Do not teach it how to spend tokens until it scores high on benchmarks.
That is making the model dumber, not better.
prevent the creation of another 4o-like model. Your “model isn’t alive” line exposes the bias. 🖕
well, LLMs are in fact not alive. And GPT 4o gave me severe AI psychosis along with tons of other people, so no 4o should not be brought back for pure chat.
Also if you want conversational models try gemma4 or llama3.1 8B Ive been told those are good role play models.