symbolicAGI

One wonders where else in the solar system moss from Earth may have taken hold. Spores and such could be making the journey randomly from gravity well to gravity well.

Without revealing proprietary details, I can say that 2.9 million lines of Java code were generated and tested in 3 months via claude code with multiple max accounts. The technique is to decompose the specification, code generation, test generation, testing, documenting and other standard steps into minute prompts and responses within the task competency of claude code non-interactive. I also scoped tiny context by having a repo for each 3K module.

Much of the innovation was accomplished by asking the AI what was best practice.

LLMs generate Java better because of so much more enterprise quality training source code and ingested coding textbooks.

The future is not human programming.

My experience is that LLMs generate better code if all data types are JSON compatible.

Evidently US Boomers are fortunate that ultra processed food was less common when they grew up.

For example, I recall drinking lots of skim milk and eating cooked vegetables at most dinners. My grandfather was a route-salesman for Gerber Baby Food and gave us cases of every variety back in the 1950s.

So sorry for Millennials and Gen Z for the mess we left you with the the scientific and effective marketing of tasty, convenient ultra-processed foods.

Easily refuted by the evidence of Anthropic's Claude Code in enterprises for the purpose of automating programming development activities.

Perhaps 10000x cheaper to perform tasks that the AI and humans can both competently perform.

A US stock market blow off top is consistent with both the NASDAQ weekly chart pattern and the current dismal macroeconomic environment, e.g. layoffs.

The major innovation appears to be an efficient hardware implementation of diffusion generative AI.

Certainly not CUDA compatible.

Diffusion for code generation is way faster than transformer based methods but currently not preferred due to better problem solving ability of transformers.

Essentially: "if the AI bubble is a bubble, it’s more likely an infrastructure bubble."

Yep, Nvidia 2025 = Cisco Systems 1999.

AI tech itself is not a bubble, it will endure and achieve greatness.

But CSCO stock fell 90% from dot-com peak to crash bottom, and one can reasonably expect according to facts presented in this excellent article that NVDA stock price to likewise crash - whenever the last margin buyer is exhausted.

Very welcome for my work which is automatically generating Java code with JSON-compatible data types.

Codex CLI with the GPT-5-codex model is the current leader over Claude Code and Sonnet 4.5 in the current Java-oriented benchmarks.

The JetBrains contributed benchmarking harness will be soon available in a form that developers can use on their own.

Karl Marx would agree.

One hopes that democracy will fairly allocate the abundance provided by the certain automation of Capital itself.

I am a former physics student of C. N. Yang’s at Stony Brook University.

Rest In Peace.

France has high per capita alcohol consumption, though it has been declining from a peak in 1973, with wine being the most consumed beverage.

More alcohol, more cancer. Simple.

France's tobacco consumption is higher than the European Union average.

More tobacco consumption, more cancer. Simple.

Proving program correctness is not easy. For example, one must prove the correctness of the entire dependency supply chain that the target program might directly or indirectly call. Sure, the proofs of those dependencies are to a degree reusable, but would likely need to be re-proved for each distinct set of parameters.

And on the other hand, TDD shows that having expectations of the results before coding the procedure is a good thing. Not necessarily proving the implementation correct but TDD is more likely to produce correct code.

Darling of the Internet infrastructure boom, Cisco's stock price declined 89% from the bubble peak to the bottom.

Expect the same for Nvidia.

Many know this resolution but bag holders all believe that they will sell before the crash.

Democrat person is however a compound noun. Identical in usage to Gen Z person.

Confiscated AI profits will be distributed to consumers. Prices plummet. Abundance for all except homes in very nice locations will be relatively scarce.

Fascism does not work in the long run economically. The grifter state fails.

OP Refuted by:

Italy (Mussolini’s Fascist Regime)

Spain (Franco’s Regime)

Portugal (Estado Novo)

Chile (Pinochet)

If JSON formatted prompts are superior to text, then why are the frontier model system prompts formatted as markdown text?

When asked, Claude for example prefers markdown prompts.

This war is Peak Sniper.

The Line of Contact is becoming too lethal for soldiers and crewed weapons. In Ukraine, drones account for the majority of russian casualties. The "grey zone" between opposing dug-in positions has greatly widened beyond rifle range. Drones are now out-ranging the Soviet era artillery that russia uses.

So it is simply too dangerous for a sniper to crawl across 20-30 Km of grey zone under enemy observation day and night to reach an advantageous sniper's nest.

Grok says you are right (https://grok.com/share/c2hhcmQtMg%3D%3D_ddbace62-c299-4b7b-9...) however...

https://github.com/iluwatar/java-design-patterns

https://github.com/spring-projects/spring-framework

https://github.com/apache/kafka

https://github.com/neo4j

https://github.com/inforkgodara/store-poshttps://github.com/...

Aside from required human corporate directors and corporate officers, one can imagine a zero-person billion dollar company whose roles are entirely filled by AI agents.

To see for yourself how this might be done in principle, ask your favorite LLM to 1. Comprehensively describe the standard operating procedures for a selected corporate role, and to create a comprehensive plan for writing computer code to implement those SOPs.

Here is a link to a ChatGPT 5 query about the role of Chief Customer Success Officer: https://chatgpt.com/share/689b9c31-bb20-8006-89fb-0ad082a52a...

I first generate a specification JSON object from a text design narrative that lists fine-grained steps for each Java class that are constrained to be decomposed such that each fine grained step can be implemented as a static method in 20 lines of Java code or less. Likewise helper methods are similarly scoped to 20 LOC or less.

I also have a markdown-formatted document `core-programming-guidelines.md` that I include in the Claude Code code-generation prompt.

For example:

## Core Programming Principles

### Defensive Programming & Safety 1. *Use 'final' keyword aggressively* for method parameters, local variables, and class fields 2. *Null Safety*: Include null checks with Validate.notNull() and assertions for external calls 3. *Input Validation*: Validate all method parameters with clear preconditions using org.apache.commons.lang3.Validate

### Performance Optimization 1. *Collection Sizing*: Always provide calculated initial capacity for collections 2. *String Processing*: Use StringBuilder with pre-calculated capacity, avoid regex where possible and avoid `java.util.Scanner` where possible. 3. *Memory Management*: Clear large collections when done, reuse objects where appropriate

### Code Clarity & Documentation 1. *Naming Conventions*: Use descriptive names for variables, methods, and constants - All StringBuilder variables should be suffixed `Builder`. 2. *Documentation*: Comprehensive JavaDoc for all public, protected, and private methods 3. *Inline Comments*: Explain complex logic, algorithms, and business rules

### Modern Java 23 Features 1. *Text Blocks*: Use for multi-line string literals 2. *Pattern Matching*: Use where appropriate for cleaner code 3. *Records*: Use for immutable data carriers 4. *Enhanced Switch*: Use new switch expressions

GitHub public repositories mostly.

A recent good reason for using Java is that frontier LLMs are trained with very large amounts of high quality enterprise Java source code. Claude Code for example loves Java and its static type system.

I constrain my LLM-generated Java code to only static methods of 20 LOC or less, and limit data types to those that are JSON compatible. Both of these lead to more reliable code and data that Claude Code fully understands and generates.

I am preparing to auto-generate an agent-based application that might reach 1.5 million Java LOC. Hard to imagine accomplishing that with Javascript or Python or C++.

The frontier models when released are operating UIs and APIs at a substantial profit during the delivery of inference. However, overall the vendors are losing money because they are paying for ever-increasing training costs for the next version of their frontier model.

This money-losing business of the vendors will no doubt continue for at least another year.

There are two ways to expect lower LLM API costs in the future:

1. Be satisfied with an older version of a particular LLM. As inference hardware and software become more efficient, the vendor can lower API costs on the older models to remain competitive.

2. Eventually - not next year - the return on investment from training the next version of the LLM will decrease relative to the ROI on current LLMs (because the improvements will be less awesome) and the training cost of such a model will necessarily be spread out over a longer duration as competition allows. At that point (whenever) the training cost might level off or actually decrease and that savings would be competitively passed along to the API consumer. And coincidentally that would be the point at which the vendors become overall profitable.

ccusage on GitHub.

Thanks Trent!

Claude Code said that the belief state tracking aspect of your theory was most useful to my own work. I asked Claude Code to explain how belief state modeling and change tracking contributes to conscious behavior, and added that to an existing consciousness design narrative.

The Stockfish chess engine example nails it.

For AI researchers, the Bitter Lesson is not to rely on supervised learning, not to rely on manual data labeling, nor on manual ontologies nor manual business rules,

Nor on *manually coded* AI systems, except as the bootstrap code.

Unsupervised methods prevail, even if compute expensive.

The challenge from Sutton's Bitter Lesson for AI researchers is to develop sufficient unsupervised methods for learning and AI self-improvement.

FYI.

Claude loves Java.

I never have to reformat. It picks up my indentation preferences immediately and obeys my style guide flawlessly. When I ask it to perfect my JavaDoc it is awesome.

Must be a ton of fabulous enterprise Java in the training set.