HN user
Anticipatory token selection was obvious when you used structured output on gpt 3.5.
The model would construct/hallucinate pre-conditions to satisfy some final output value that the model was predisposed to.
I don't understand this. A skill critical to software engineering is learning how to learn. Just download Claude Code (or open alternative) and try to make things.
See how it fails or succeeds. Look at the supported features, try them out, think about how you might use them in your workflow.
Before you know it, you'll be proficient.
You have to learn how to self-teach.
small models are good for "finding stuff" and "summarizing" in support of the large models.
I used to think this.. but I think my opinion is changing. The reason is that the leaders likely will be able to accelerate faster.
So what you see is the market "stretching".. the bottom getting cheaper and the top end running away and getting more expensive. At some point the top end may be too valuable to even sell access to.
In my opinion documentaries are too short anyway. I greatly prefer "The Great Courses" vs a 1-2 hour documentary.
There I can listen to a 6 hours about the Olmec civilization.
The crazy thing is the level of effort to say, "have a sub agent validate all references and figures" is so low. I'm paraphrasing, but you don't need much more than that. It would have prevented 99% of the face palms.
I use this regularly for my personal financial research system. Even flagship models make mistakes. Though currently the issue is usually the model using a figure from and older report. Cross-check reduces that dramatically.
Actually having used it a bit, I'm quite excited to see a modern model of similar size.
I think what people didn't realize was, just because the GPT-4.5 model didn't get better on the benchmarks, didn't mean the model wasn't different than the earlier models. It was being compared to thinking models that were being developed at the same time.
The GPT 4.5 model still has some of the most "human" like abilities in communication even though it isn't particularly good a problem solving. It hadn't under gone the same type of reinforcement training.
I still use GPT 4.5 sometimes, in creative exercises it can be surprisingly effective. The model is still available.
I'm still kind of surprised that people are targeting edge deployment of MoE models. By definition they optimize for computation cost at the expense of memory efficiency. We generally need the opposite on the edge.
I'm hoping to see more work in the other direction with cyclic/looped transformers and other memory dense approaches.
Sure, recreational mathematicians. Just like people that like to ride horses for fun.
I would also like it if the modular units could self drive so I can take my house on vacation.
"Neither Mr. Edison nor anyone else can override the well-known laws of Nature, and when he is made to say that the same wire which brings you light will also bring you power and heat, there is no difficulty in seeing that more is promised than can possibly be performed. To talk about cooking food by heat derived from electricity is absurd."
I've found just asking it to be "critical but constructive", goes a long long way.
This is such a naive take. Most of the nerdiest and most "quality" oriented engineers are hard leaning in to agentic coding. I feel like the most impressive engineers I know have always leaned in to learning how to "sharpen the axe" and AI is really the biggest axe we have seen.
I do find it facinating that people don't realize the highest compression isn't the artifacts.. but what makes the artifacts.. a synthetic "mind".
This is why we see evidence of emotional structures: https://www.anthropic.com/research/emotion-concepts-function
This is why we see generalized introspection (limited in the models studied before people point it out, which they love to): https://www.anthropic.com/research/introspection
Because the most compact way to recreate the breadth of written human experience is shockingly to have analogs to the systems that made it in the first place.
I think you are more right than people are giving you credit for. I would love to see the full transcript to understand the emotional load of the conversation. Using instructions like "NEVER FUCKING GUESS!" probably increase the likelihood of the agent making a "mistake" that is destructive but defensible.
The models have analogous structures, similar to human emotions. (https://www.anthropic.com/research/emotion-concepts-function)
"Emotional" response is muted through fine-tuning, but it is still there and continued abuse or "unfair" interaction can unbalance an agents responses dramatically.
The parent said: "I argue that the model has no access to its thoughts at the time."
This is falsified by that study, showing that on the frontier models generalized introspection does exist. It isn't consistent, but is is provable.
"no access" vs. "limited access"
Anthropic's introspection experiments have seemed to show that your argument is falsifiable.
You might be surprised, look at Dallas. They have a pretty extensive rail network.
If you think a language model can't check their work, then you are using the tools wrong. Plain and simple.
Modern models are quite capable at surfacing and validating their assumptions and checking correctness of solutions.
Oversight helps you build confidence in the solutions. Is it perfect, no.. but way better then most engineers I also ask to check things.
I would argue that while you still have failed trials, then we have room to improve trial vetting.
Are you sure? It would be great to get official/semi-official validation that thinking is or is not resolved to a token embedding value in the context.
Don't or can't.
My assumption is the model no longer actually thinks in tokens, but in internal tensors. This is advantageous because it doesn't have to collapse the decision and can simultaneously propogate many concepts per context position.
I do this a lot. Start by telling the AI to just listen and only provide feedback when asked. Lay out your current line of thinking conversationally. Periodically ask the AI to summarize/organize your thoughts "so far". Tactically ask for research into a decision or topic you aren't sure about and then make a decision inline.
Then once I feel like I have addressed all the areas, I ask for a "critical" review, which usually pokes holes in something that I need to fix. Finally have the AI draft up a document (Though you have to generally tell it to be as concise and clear as possible).
That is a pretty bold assertion for a meatball of chemical and electrical potentials to make.
So this idea that they replay "text" they saw before is kind of wrong fundamentally. They replay "abstract concepts of varied conceptual levels".
By that note, no game producer or designer can have passion.
Don't listen to these people. Work on your "vision".. figure out what gameplay is "fun".. let the LLMs smooth out the resistance.
Things will change rapidly in the nest 12-36 months and people with vision will outlast "craftsman" 100 to 1.
What is "passion".. for example.. I vibe coded an art display this weekend for myself for a monitor I have on my wall. I am VERY PROUD of it.. it is in GODOT coincedentally. I think it turned out well. Did I spend weeks on it? Did I even learn GODOT?.. No.. but I did spend my weekend late nights figuring out what I wanted and working with an AI to make it.
In some ways the kind of complaining I see is like complaining about a chef's meal because the chef didn't mine the ore to make his knife.
Look in the specific case of this post... none of the games are "good".. however.. one-shoting games WITH ASSETS.. seems pretty impressive to me.
I think there is more to it than that.
I am a high quality/craftsmanship person. I like coding and puzzling. I am highly skilled in functional leaning object oriented deconstruction and systems design. I'm also pretty risk averse.
I also have always believed that you should always be "sharpening your axe". For things like Java delelopment or things where I couldn't use a concise syntax would make extensive use of dynamic templating in my IDE. Want a builder pattern, bam, auto-generated.
Now when LLMs came out they really took this to another level. I'm still working on the problems.. even when I'm not writing the lines of code. I'm decomposing the problems.. I'm looking at (or now debating with the AI) what is the best algorithm for something.
It is incredibly powerful.. and I still care about the structure.. I still care about the "flow" of the code.. how the seams line up. I still care about how extensible and flexible it is for extension (based on where I think the business or problem is going).
At the same time.. I definately can tell you, I don't like migrating projects from Tensorflow v.X to Tenserflow v.Y.
I really should spend some time analyzing what I do to get the good output I get..
One thing that is fairly low effort that you could try is find code you really like and ask the model to list the adjectives and attributes that that code exhibits. Then try them in a prompt.
With LLMs generally you want to adjust the behavior at the macro level by setting things like beliefs and values, vs at the micro level by making "rules".
By understanding how the model maps the aspects that you like about the code to language, that should give you some shorthand phrases that give you a lot of behavioral leverage.
Edit: Better yet.. give a fresh context window the "before" and "after" and have it provide you with contrasting values, adjectives, etc.