Support NeoGAF

[IbizaPocholo]

https://blog.e01.ai/glm5-gameboy-and-long-task-era-64db7074a026

GLM5 excels in long-task engineering by maintaining consistency over 700+ tool calls and 800+ context handoffs across 24+ hours, enabling AI to function as a persistent process for complex projects like building a full Game Boy Advance emulator, rather than short conversational interactions.

Summary

• E01 Research gained early access to GLM5 to test its long-task capabilities, focusing on a challenge to build a Game Boy Advance (GBA) emulator from scratch in JavaScript, embedded in a 3D rendered scene, using a single agent without parallelism.
• The task simulates real engineering: research, architecture, implementation, testing, and documentation across multiple sessions, stretching beyond context limits via meta-rules and loops.
• Challenge input: system prompt and hardware documentation; model must scope work, adjust strategies, switch roles (architect, engineer, designer), and hand off context accurately.
• Two test versions: "Easy mode" with gbajs reference code (GLM5 reimplemented independently, achieving working core emulator, ROM loading, 3D scene; demo at https://e01.ai/gba); "Zero reference" mode (no code or web search, ran 24+ hours, completed CPU instruction set, ongoing progress).
• Prior models failed by looping, forgetting goals, or erroneous tool calls.
• Success mechanism: Prompt as a meta-loop (work → test → log → advance), persisting state in files (/notes/progress.md, /notes/decisions.md, /notes/blockers.md) for context resets.
• Observation 1: GLM5 showed no degradation—consistent tool calls (700+), strict instruction adherence (800+ switches), reliable context relay from files.
• Implications: Enables goal-driven agents (autonomous planning/execution), parallel multi-agents (one human supervising many), applications beyond code (e.g., AI for Science: experiment design, research); patterns like long-recurring (iterative workflows) and long-exploring (open-ended exploration).
• Observation 2: Challenges include hidden multi-session loops (human-detected, e.g., brute-forcing 3D model), over-diligence without pause thresholds; needs explicit "stop and ask for help" instructions.
• Future needs: Observability (visualization/monitoring), intervention (alerts, nudges), evaluation metrics (context relay, progression rate, decay), trust (incremental validation), cost/infra (budgeting, pause/resume), research (relay limits, self-evaluation).
• Prompt design guide: Define goal + phases with "done" criteria (e.g., CPU core, memory, graphics); conventions (file structure, testing); notes protocol (session updates); testing gates (unit/integration via Node.js, not browser); loop breaking (retry logs, time limits); recovery (read notes/files first).
• Mistakes to avoid: Vague notes, no loop-breaking/logs (resets across sessions), assuming memory persistence, over-specifying code, skipping tests (compounds errors).
• Experiment run via OpenCode/Claude Code; tags include GLM5, AI, Long Task, Writing Prompts, Machine Learning.

 

[yurinka]

wtf, wow

 

[Faltherion]

This is absolutely crazy.

 

[bighugeguns]

Keep being told how useless AI is then suddenly the craziest shit you ever saw is shown, there's a small pause, and you keep getting told AI is useless.

 

[peish]

It's just regurgitating data from existing emulators, don't trust these ai bots

 

[Biff]

• Two test versions: "Easy mode" with gbajs reference code (GLM5 reimplemented independently, achieving working core emulator, ROM loading, 3D scene; demo at https://e01.ai/gba); "Zero reference" mode (no code or web search, ran 24+ hours, completed CPU instruction set, ongoing progress).

So... it had access to the complete code of a working GBA emulator? Is "reimplemented" just a fancy way of saying "copied", lmao?

And the version of the model that had no existing GBA reference code ran for 24 hours straight and wasn't able to get it working. "Completed CPU instruction set" wtf does that even mean.

I want AGI as badly as the next guy, but I also equally want to avoid getting duped by people who are heavily incentivized to make us believe AGI is around the corner.

 

[Spukc]

 

[March Climber]

Keep being told how useless AI is then suddenly the craziest shit you ever saw is shown, there's a small pause, and you keep getting told AI is useless.

It's a great tool.

A lot of tools can be useless depending on who uses it.

 

[kiphalfton]

An emulator creating an emulator.

 

[Autistic_Pancake]

It's just regurgitating data from existing emulators, don't trust these ai bots

Only if the said data was present in its training dataset, and if the said dataset was built well enough not to lose any important contents.

On a side note, GLM 5 is twice as larger (= 2x increased VRAM or RAM+VRAM requirements) than GLM 4.7, despite only being 10 to 15% better in benchmarks.

Even though the advances are there, they come at a ridiculous cost and there's no optimization (that doesn't rely on lobotomizing the model) in sight...

 

[SolarFry]

AI marketing in overdrive, it's attacking from everywhere. I know companies are pouring 100s of billions in it but it's getting tiresome.

 

[PonyStation4]

AI generates GBA emulator by stealing code from existing GBA emulators. Very cool to see

 

[StueyDuck]

AI marketing in overdrive, it's attacking from everywhere. I know companies are pouring 100s of billions in it but it's getting tiresome.

Oh we are gonna get a million "this ai is actually sentient, oh wait no it isnt" marketing "scandals" from all these snake oil salesmen.

I love AI, use it daily for work and personal use. The salesman are indeed super insufferable

 

[Jinzo Prime]

At first I was like:

Couple posts later I'm like:

Headlines are fun.

 

[Dutchy]

Was about to be impressed by something-AI for the very first time but then saw it was an Ibiza thread so there had to be a catch.

 

[I wasn't selling any]

Keep being told how useless AI is then suddenly the craziest shit you ever saw is shown, there's a small pause, and you keep getting told AI is useless.

You consider simply rewriting other peoples (humans) work/code to be "the craziest shit you ever saw"? AI usage can be very helpful, but this particular case shows that's it's really useless most of the time.

 

[MidGenRefresh]

Keep being told how useless AI is then suddenly the craziest shit you ever saw is shown, there's a small pause, and you keep getting told AI is useless.

Crazy shit would be if it could write a fully working Switch 2 or PS5 emulator from scratch. But it won't. It can't truly create something entirely new that wasn't, in some form, already reflected in its training data.

 

[bighugeguns]

Crazy shit would be if it could write a fully working Switch 2 or PS5 emulator from scratch. But it won't. It can't truly create something entirely new that wasn't, in some form, already reflected in its training data.

This is a good explanation. You're right.

 

[Knightime_X]

It's just regurgitating data from existing emulators, don't trust these ai bots

Don't look at ai where it currently stands, look at where its heading.
ai works like a person, eventually it won't need people for any kind of help.
A person doesn't know how to draw a circle or know what 1+1 is without another human's help.
Ai basically functions the same way, only vastly superior as it learns.

 

[Guilty_AI]

AI news cycle be like:

>AI did some crazy shit
>"Wow, thats incredible! Take that AI deniers!"
>There's a very wonder-breaking catch
>"Nono, the future is what matters! It'll be great in the next 6 months i swear!"
>RAM prices increase by 14278%

 

[Guilty_AI]

On a more serious note, the conclusion i'm getting at is that these bots, at the end of the day, will work better as assistants for very specific things. Mainly for repetitive tasks and other types of mule jobs, doesn't seem they'll ever be able to produce something (good) on their own.

 

[peish]

Don't look at ai where it currently stands, look at where its heading.
ai works like a person, eventually it won't need people for any kind of help.
A person doesn't know how to draw a circle or know what 1+1 is without another human's help.
Ai basically functions the same way, only vastly superior as it learns.

Doubt it! They must have been fed gigazillionbytes of data by now, guzzled up megawattons of electricity by now. Still waiting for Ai to learn!

 

[nowhat]

Don't look at ai where it currently stands, look at where its heading.

Where it is heading? Seems to be going around in circles as far as I can tell. Although that image is somewhat outdated, as investments in OpenAI seem to be drying up - and that companies are moving away from NVidia when it comes to inference. Which should make for a rather interesting stock market in the near future.

 

[Knightime_X]

Where it is heading? Seems to be going around in circles as far as I can tell. Although that image is somewhat outdated, as investments in OpenAI seem to be drying up - and that companies are moving away from NVidia when it comes to inference. Which should make for a rather interesting stock market in the near future.

Ask me again in a few years.
I remember hearing the same thing about cell phones and the internet.

 

[Guilty_AI]

Ask me again in a few years.
I remember hearing the same thing about cell phones and the internet.

Wasn't the .com bubble burst one of the most famous ones? AI seems to heading on a similar path, where the tech hype for it doesn't match it's actual uses (that we'll probably only truly understand a decade or so from now).

 

[cireza]

In today's news, AI invented copy / paste. Was definitely worth the investment.

 

[poppabk]

Wasn't the .com bubble burst one of the most famous ones? AI seems to heading on a similar path, where the tech hype for it doesn't match it's actual uses (that we'll probably only truly understand a decade or so from now).

But all those uses became a reality - people just jumped the gun - invested too much too fast. But that was mainly venture capital - that once it was gone it was gone. With AI it is big companies with deep pockets who will likely ride the dips in the hopes that they end up with an Amazon.com and not a boo.com.

 

[Lukaku's First Touch]

Sure it might process this but at what cost? The power / water / etc... costs will be huge. Also they'll just make your company dependent on AI and then start charging you giant fees to use their AI that might even be more expensive than some engineers.

 

[john2gr]

It would be more impressive if they used AI to create a perfect Dreamcast emulator, or a Model 3/NAOMI arcade emulator. I know there are DC and Model 3 emulators out there, but they are not perfect. So, AI filling the gaps and improving them would be a cool concept.

Or, for the ultimate challenge, a PS4 emulator that can run all PS4 games without any graphical glitches.