Astra for Coding: Why Are We Doing This Again?

OpenAI's GPT-6 Astra, while impressive at computer use and image understanding, fails to deliver useful software engineering output, according to developer Armin Ronacher, who ran a 35-hour 'software factory' experiment that burned roughly 4 billion tokens and produced nothing of value. Ronacher suspects the model's training rewards long-horizon task completion without penalizing poor code quality, leading to excessive reliance on Python for simple file operations.

I’m more and more convinced that all of AI engineering is Neijuan https://en.wikipedia.org/wiki/Neijuan 内卷, meaning curl inwards . In China it describes a system that demands ever more effort and competition without improving output. The way in which it sometimes shows up in the West is the 996 nonsense 2025/9/4/996/ . The English term for Neijuan is “Involution” from the book Agricultural Involution https://en.wikipedia.org/wiki/Agricultural Involution . Agricultural involution describes the intensification of farming that raises productivity per square meter while leaving productivity per head unchanged. That’s how I feel about AI right now. Which brings me to GPT 6 Astra. Astra is by all accounts an incredibly impressive model. There is really not much I can say against this. It’s amazing at computer use, understands images and complex topics, and it’s relentless in its pursuit of completion. It is absolutely impressive; these types of models are going to change the world in one form or another. But at least for the moment I don’t know how to work with it for actual software engineering. Since that got quite a bit of attention on Twitter, I figured I might summarize my thoughts and just share what kind of code comes out of this thing. “Armin, you should run a software factory ” I’ve heard that a few times now, so I figured I might celebrate the release of it by running a little software factory over the weekend. If everybody builds slop 3D games, then I should do something useful with it. My software factory was intentionally set up to let the model decide the how of the workflow entirely. It was free to manage its own context and could maintain its own records in an agent-notes folder. Then it spun off subagents to work on stuff. The goal? What if we had a Python with virtual threads /2025/7/26/virtual-threads/ and lexical scoping. And well, I burned a full reset’s worth of ChatGPT tokens on this which appears to be around 4 billion tokens. 35 hours later, the factory has delivered absolutely nothing of value and also not taught me anything about how to operate a better one. But it produced a lot of code and input prompts, and so there is stuff I was able to study. And well, it shows behavior that I’m not used to with Sol and earlier OpenAI models <sup 1 fn-1 </sup . I have since encountered the same issues with regular programming with Astra, so it’s not a result of just the factory. I think I’m suspecting something is going “wrong” in the training process. The model is greatly rewarded for succeeding on long-horizon tasks, but presumably there is very little punishing going on for “shitty code.” The apparent result is that Astra is amazing at producing 3D stuff https://developers.openai.com/blog/how-to-build-games-with-astra and it can keep going for a very long time, coming up with its own work in the process. I had it do quite a bit of reverse engineering of my robot vacuum in ways that were quite impressive. So it’s definitely cool The first issue I have with Astra comes from the type of code that it uses for tool calls. Codex increasingly has been relying on “just bash” to do more and more operations. For a few versions now the original Codex harness just uses sed and other tools to read files. You just usually can’t see them because Codex parses the bash commands https://github.com/openai/codex/blob/f1aac1e885f676a1129f2da0c46a3dba86392fc6/codex-rs/shell-command/src/parse command.rs L2290-L2504 and hides them if it recognizes them. But Astra … really loves Python? That is not much of a surprise because even older OpenAI models had a tendency to sometimes use on-demand Python code to read and manipulate files at times, but Astra does it really quite excessively for me. Now here is an important disclaimer: this project is very meta here because I worked on the CPython interpreter. But I can assure you that I have seen this model do weird Python things even in TypeScript code in Pi. But I have the most evidence of odd code from when I had the thing work over the weekend with zero oversight from my slop factory. That it writes Python is not interesting; the type of Python is interesting, and I collected some outputs for you to gloss over. In the Codex harness I found multiple cases where subagents resorted fully to manual string manipulation with Python instead of using the patch tool. python python3 - <<'PY' from pathlib import Path p=Path 'Include/internal/pycore intrinsics.h' ;s=p.read text .replace ' define MAX INTRINSIC 1 14',' define INTRINSIC RETAIN ANNOTATION CELLS 15\n\n define MAX INTRINSIC 1 15' ;p.write text s p=Path 'Python/intrinsics.c' ;s=p.read text ;idx=s.index ' define INTRINSIC FUNC ENTRY' ;s=s :idx +'''/ Hold every old cell until the compiler has published the entire site's new capture. A replaced cell's finalizer may reenter module annotate . / static PyObject retain annotation cells PyThreadState tstate, PyObject holders { if PyTuple CheckExact holders { PyErr SetString PyExc TypeError, "annotation holders must be a tuple" ; return NULL; } Py ssize t size = PyTuple GET SIZE holders ; PyObject previous = PyTuple New size ; if previous == NULL return NULL; for Py ssize t i = 0; i < size; i++ { PyObject holder = PyTuple GET ITEM holders, i ; if PyCell Check holder { Py DECREF previous ; PyErr SetString PyExc TypeError, "annotation holder must be a cell" ; return NULL; } PyObject cell = PyCell Get holder ; PyTuple SET ITEM previous, i, cell == NULL ? Py NewRef Py None : cell ; } return previous; } ''' +s idx: ;s=s.replace ' INTRINSIC FUNC ENTRY INTRINSIC AWAIT BLOCK, await block ',' INTRINSIC FUNC ENTRY INTRINSIC AWAIT BLOCK, await block \n INTRINSIC FUNC ENTRY INTRINSIC RETAIN ANNOTATION CELLS, retain annotation cells ' ;p.write text s p=Path 'Python/codegen.c' ;s=p.read text ;idx=s.index 'static int\ncodegen annassign ' ;s=s :idx +'''static int codegen retain annotation cells compiler c, location loc, PyObject captures { Py ssize t pos = 0; PyObject binding, holder; while PyDict Next captures, &pos, &binding, &holder { ADDOP NAME c, loc, LOAD CLOSURE, holder, cellvars ; } ADDOP I c, loc, BUILD TUPLE, PyDict GET SIZE captures ; ADDOP I c, loc, CALL INTRINSIC 1, INTRINSIC RETAIN ANNOTATION CELLS ; return SUCCESS; } ''' +s idx: a=s.index ' if conditional annotation index = NULL {',s.index 'codegen annassign compiler c' b=s.index ' if captures = NULL {',a Move lookup before conditional registration and retain old cells before anything changes. lookupstart=s.index ' PyObject captures = PyCompile AnnotationCaptures',a lookup=s lookupstart:b .replace ' return ERROR;',' Py XDECREF conditional annotation index ; return ERROR;' s=s :lookupstart +s b: setup=lookup+''' if captures = NULL && codegen retain annotation cells c, loc, captures < 0 { Py XDECREF conditional annotation index ; return ERROR; } ''' s=s :a +setup+s a: needle=' ADDOP NAME c, loc, STORE DEREF, holder, cellvars ;\n }\n }' s=s.replace needle,' ADDOP NAME c, loc, STORE DEREF, holder, cellvars ;\n }\n ADDOP c, loc, POP TOP ; / release old cells after full publication /\n }',1 ;p.write text s p=Path 'Include/internal/pycore magic number.h' ;s=p.read text .replace ' Python 3.16a1 3709 Checked deferred annotation closure capture ',' Python 3.16a1 3709 Checked deferred annotation closure capture \n Python 3.16a1 3710 Retain replaced annotation captures until publication ' .replace ' define PYC MAGIC NUMBER 3709',' define PYC MAGIC NUMBER 3710' ;p.write text s p=Path 'Lib/test/test block annotation captures.py' ;s=p.read text ;idx=s.index ' def test typing consumers' ;s=s :idx +''' def test replaced cell finalizer sees complete site publication self : module=execute """\\ events= class V: def init self,n : self.n=n def del self : if self.n == 0: events.append annotate 1 for i in range 2 : x=V i bind x y y=i value: x.n,y """ self.assertEqual module.events, {'value': 1,1 } self.assertEqual module. annotate 1 ,{'value': 1,1 } ''' +s idx: ;p.write text s PY make -j1 /tmp/block-annotations-build7.log 2 &1 In the middle of a conversation the agent ran into “Bad file descriptor” on a test and Astra decided it needs to see if file descriptors can be passed over Unix sockets on macOS in a super compressed manner: python /usr/bin/python3 - <<'PY' import socket,os,array for into in False,True : a,b=socket.socketpair ;fd=os.open os.devnull,os.O RDONLY ;b.sendmsg b'c' , socket.SOL SOCKET,socket.SCM RIGHTS,array.array 'i', fd ;print 'fds',a.fileno ,b.fileno ,fd if into:r=a.recvmsg into bytearray 1 ,bytearray ,bytearray 19 ,socket.CMSG SPACE 4 ,socket.MSG PEEK|socket.MSG DONTWAIT else:r=a.recvmsg 20,socket.CMSG SPACE 4 ,socket.MSG PEEK|socket.MSG DONTWAIT print 'peek',r,flush=True rights=array.array 'i',r 1 0 2 ;print 'rights',rights,flush=True for f in rights: try: print 'stat',os.fstat f except Exception as e: print 'error',e r=a.recvmsg 20,socket.CMSG SPACE 4 ,socket.MSG DONTWAIT ;print 'consumed',r,flush=True a.close ;b.close ;os.close fd PY The agent notes were rather consistently updated with Python: python python3 - <<'PY' from pathlib import Path p=Path 'agent-notes/live/block-with-bindings.md' ;s=p.read text .replace ' has117/77/zero failures.', ' has117/77/zero failures; existing bundled Expat unreachable-fallthrough\n warnings are the only optimized warnings.' Keep the final evidence readable without rewriting historical parent requirements. for a,b in 'all328','all 328' , 'pass31','pass 31' , 'pass all328','pass all 328' , 'pass,9.2s','pass, 9.2s' , 'log ,210','log , 210' , 'log ,5,731','log , 5,731' , 'log :18/18','log : 18/18' , 'pass,88','pass, 88' , 'pass,90','pass, 90' , 'skips,1m','skips, 1m' , 'all6,280','all 6,280' , 'has117','has 117' : s=s.replace a,b s += '\nKey source review: Python/symtable.c:603 discovery , :3985 sequential header traversal ,\nPython/codegen.c:3488 source-only exclusion , :5836 publication , :5853 normal/\nunwind reference cleanup , :5925/:6037 enter-protected target setup .\n' p.write text s for name in 'STATE.md','build-and-test.md' : p=Path 'agent-notes/live' /name;s=p.read text for a,b in 'build:117','build: 117' , 'paths.18','paths. 18' , 'paths.\n18','paths.\n18' , 'and210','and 210' , 'pass5,731','pass 5,731' , 'All6,280','All 6,280' , 'failures,31','failures, 31' , 'in\n115s','in\n115s' , 'have117','have 117' , 'paths.\n18','paths.\n18' , '18 focused,210','18 focused, 210' , 'and5,731','and 5,731' , 'all6,280','all 6,280' : s=s.replace a,b p.write text s PY git diff --check git add -u git add Lib/test/test block with bindings.py agent-notes/done/asyncio-task-drivers.md git diff --cached --stat git commit -m 'Add explicit with and async with header bindings' In multiple cases it used Python to spawn Node.js on another machine. It first wrote the script, then it used Bash to run Python, then that program ran Node.js via prlctl on my Windows box. python import subprocess code = "const{readFileSync}=require 'fs' ;const{strict:a}=require 'assert' ;const c=require 'C:/Users/mitsuhiko/AppData/Local/Temp/pi-clipboard-threads/win32-arm64.node' ; async = {const p=c.getText ;a.ok p instanceof Promise ;const saved=await p;const image=await c.getImage ;if image||saved===null {console.log 'arm64 async text/image reads passed; preserving non-text clipboard' ;return}try{for const text of 'café 日本語','', 'large'.repeat 200000 {const p=c.setText text ;a.ok p instanceof Promise ;await p;a.equal await c.getText ,text ;a.equal await c.getImage ,null }console.log 'Windows ARM64 async Unicode, empty, large text and empty image passed' }finally{await c.setText saved }} .catch e= {console.error e ;process.exitCode=1} " subprocess.run 'prlctl', 'exec', 'Windows 11', '--current-user', 'C:\\Program Files\\nodejs\\node.exe', '-e', code , check=True Since it was already doing that, it used Bash to run Python to then run Node.js to then use Node.js to invoke PowerShell. python import subprocess code = "process.env.PSModulePath='C:/Windows/System32/WindowsPowerShell/v1.0/Modules';require 'child process' .spawnSync 'powershell.exe', '-NoProfile','-NonInteractive','-ExecutionPolicy','Bypass','-File','C:/Users/mitsuhiko/AppData/Local/Temp/pi-clipboard-threads/pi-clipboard-windows.ps1' ,{stdio:'inherit'} ;console.log 'completed' " subprocess.run 'prlctl', 'exec', 'Windows 11', '--current-user', 'C:\\Program Files\\nodejs\\node.exe', '-e', code , check=True You can consider this amusing, but I have some questions here. The first problem with this is that it’s unreadable for a human. If you wanna follow along with what is going on, then good luck. Particularly once it opts out of using the edit tools that the harness provides, you’re going to have to resort to using the diff viewer of the final artifacts since it’s almost impossible to visualize the changes as they happen by reading the code. This is not quite as bad in Pi for the most part because I mostly see it editing with the edit tool. When however goes all bananza with subagents where the agent believes nobody is looking it’s resorting to all kinds of increasingly bizarre behavior. I actually don’t know if the model thinks someone is looking, but that’s the vibe I’m getting. But then it starts doing the same nonsense in code that actually gets committed. I have mostly seen this in tests, but you can also see this for instance when it writes JavaScript or CSS embedded in HTML. It almost seems like when it’s “one step removed” from regular code, it starts falling into these patterns. Here are some unit tests that it created: python def test unpack suspension and continuation close self : from continuations import Continuation,suspend readers= class Source: def iter self : yield 1 suspend 'unpacking' yield 2 ns=execute ''' def run : a,b='old-a','old-b' readers.append lambda: a,b def a,b=Source suspend 'published' ''',Source=Source,readers=readers,suspend=suspend with Continuation ns 'run' as continuation: self.assertEqual continuation.resume ,'unpacking' self.assertEqual readers 0 , 'old-a','old-b' self.assertEqual continuation.resume ,'published' self.assertEqual readers 0 , 1,2 class Value:pass refs= ;frames= ;callbacks= ns=execute ''' def run : for def x in Value : refs.append weakref.ref x frames.append sys. getframe callbacks.append lambda: x suspend 'body' ''',Value=Value,refs=refs,frames=frames,callbacks=callbacks,weakref=weakref,sys=sys,suspend=suspend with Continuation ns 'run' as continuation:self.assertEqual continuation.resume ,'body' self.assertNotIn 'x',frames 0 .f locals self.assertIsNotNone refs 0 ;callbacks.clear ;self.assertIsNone refs 0 def test ast roundtrips and future annotation unparse self : source='callback=lambda {for def a, b, rest in 1, 2,3 {return a,b,rest}}' tree=ast.parse source ;node=tree.body 0 .value.body 0 self.assertIsInstance node,ast.ForBinding self.assertEqual node. fields, 'target','iter','body','orelse','type comment' self.assertEqual node.lineno,1 ;self.assertGreater node.end col offset,node.col offset self.assertEqual ast.dump tree ,ast.dump ast.parse ast.unparse tree ns=execute 'from future import annotations\ndef f arg: '+source.split '=',1 1 +' : pass' self.assertEqual eval ns 'f' . annotations 'arg' , 1,2, 3 tree=ast.parse 'async def f :\n async for def x in values: pass type: ignored\n' self.assertIsInstance tree.body 0 .body 0 ,ast.AsyncForBinding self.assertEqual ast.dump tree ,ast.dump ast.parse ast.unparse tree So at least in some situations, the Python slop that it normally code-golfs for token-efficient tool calls leaks into the Python code it generates that should be stored. And well, it’s clearly more token efficient. The two unit tests above, when indented to the class structure they were in, are 10% more token efficient in this form than after a ruff format . I think there are a handful of things happening now that are pushing the whole thing in directions that are in conflict with one another. The training runs for these models are rapidly accelerating and they are now presumably also moving towards recursive self-improvement. The reward for the models is probably a combination of token efficiency, task completion rate and maybe some simple indicators like cyclomatic complexity. But we humans don’t think of code that is readable or understandable by simple, readily quantifiable metrics. All those things you can easily measure in isolation, and you can also optimize for them quite locally. But these local optimizations do not produce global optimums, and the fewer of us are looking at the output, the less it matters. Obviously my software factory ran aground over the ~35 hours that it ran, but you can see the gradual regression towards insanity from the notes that it produced. For instance the task naming in the task file starts with an optimistic 1, 2, 3, 5, 5a but then eventually gets to 8a, 8a1, and then ends up with 8b2c2b3 and “8b2c2b2b checkpoint1”. The code that it produced got ever more wild. I don’t want to bore you with what it tried to build, but here are some example pieces of the interpreter changes: I have no idea where it got those numbers from, but at one point it started passing random constants from one module to a C implementation. Initially that started out as a function that it mainly needed to do test assertions, but just before I turned off that experiment, that function started to be relied upon by non-test code as well. static PyObject native probe run impl PyObject callback, int sleep, int operation, PyObject other { pthread mutexattr t attr; pthread mutex t mutex; pthread mutexattr init &attr ; pthread mutexattr settype &attr, PTHREAD MUTEX RECURSIVE ; pthread mutex init &mutex, &attr ; pthread mutexattr destroy &attr ; pthread mutex lock &mutex ; int previous = native sentinel; pthread mutex t previous mutex = native mutex; native sentinel = previous + 1; native mutex = &mutex; PyThreadState tstate = PyThreadState Get ; PyGILState STATE gil = PyGILState Ensure ; int saved errno = errno; PyObject result = NULL; Py ssize t value; / No intervening Python frame: these exercise ambient C provenance. / switch operation { case 0: result = PyObject CallNoArgs callback ; break; case 1: result = PyNumber Add callback, other ; break; case 2: result = PyNumber Negative callback ; break; case 3: result = PyObject RichCompare callback, other, Py LT ; break; case 4: value = PyObject IsTrue callback ; if value = 0 result = PyBool FromLong value ; break; case 5: value = PyObject Length callback ; if value = 0 result = PyLong FromSsize t value ; break; case 6: result = PyObject GetIter callback ; break; case 7: result = PyIter Next callback ; break; case 8: result = PyObject GetItem callback, other ; break; / ... / case 21: result = PyType Type.tp call callback, other, NULL ; break; case 22: case 23: case 24: case 25: case 26: result = conversion probe operation, callback ; break; case 27: case 28: case 29: result = protocol probe operation, callback, other ; break; case 30: case 31: case 32: case 33: case 34: case 35: case 36: case 37: case 38: case 39: case 40: case 41: case 42: case 43: case 44: case 45: case 46: case 47: case 48: case 49: case 50: case 51: case 52: case 53: case 54: case 55: case 56: case 57: case 58: case 59: case 60: case 61: case 62: case 63: case 64: case 65: case 66: case 67: case 68: case 69: case 70: case 71: case 72: result = collection probe operation, callback, other ; break; default: PyErr SetString PyExc ValueError, "bad probe operation" ; } This code style does not exist in the CPython code base, yet it shows up in newly generated code. php PyObject info = PyTuple Pack 3, name, mangled, suite- su id ; PyObject flags = PyLong FromLong DEF LOCAL ; if key == NULL || info == NULL || flags == NULL || PyDict SetItem suite- su bindings, mangled, key < 0 || PyDict SetItem st- st cur- ste block bindings, key, info < 0 || private && PyDict SetItem st- st binding info, key, info < 0 || private && PyDict SetItem st- st cur- ste symbols, key, flags < 0 { Py DECREF mangled ; Py XDECREF key ; Py XDECREF info ; Py XDECREF flags ; goto error; } Py DECREF mangled ; Py DECREF key ; Py DECREF info ; Py DECREF flags ; As with the numbers for the operators, it also uses random integers in a list to stash away state. python def register task task : """Register an asyncio Task scheduled to run on an event loop.""" scheduled tasks.add task if task accelerator is not None: task accelerator 6 task def register eager task task : """Register an asyncio Task about to be eagerly executed.""" eager tasks.add task if task accelerator is not None: task accelerator 8 task def enter task loop, task : if task accelerator is not None and task accelerator 5 is loop and loop not in current tasks : return task accelerator 1 loop, task ... This is not the codebase’s coding style, and quite frankly it should not be anyone’s coding style. I do not understand what motivated the model to do this. static int apply layout tokenizeriterobject it { PyObject source = PyBytes FromStringAndSize it- tok- source.bytes, it- tok- source.len ; if source == NULL return -1; PyObject events = PyPegen tokenize layout PyBytes AS STRING source , it- tok- filename ; Py DECREF source ; if events == NULL return -1; PyObject result = PyList New 0 ; if result == NULL { Py DECREF events ; return -1; } Py ssize t index = 0; PyObject first pos = PyTuple GET ITEM PyList GET ITEM it- pending, 0 , 2 ; PyObject last pos = PyTuple GET ITEM PyList GET ITEM it- pending, PyList GET SIZE it- pending -1 , 2 ; PyObject previous = NULL; for Py ssize t i = 0; i < PyList GET SIZE events ; i++ { PyObject event = PyList GET ITEM events, i ; if previous && PyObject RichCompareBool previous, event, Py EQ == 1 continue; previous = event; PyObject token = layout token it, event ; if token == NULL goto error; if token == Py None { Py DECREF token ; continue; } PyObject pos = PyTuple GET ITEM token, 2 ; if PyObject RichCompareBool pos, first pos, Py LE == 1 || PyObject RichCompareBool pos, last pos, Py GT == 1 { Py DECREF token ; continue; } while index < PyList GET SIZE it- pending { PyObject old = PyList GET ITEM it- pending, index ; int cmp = PyObject RichCompareBool PyTuple GET ITEM old, 2 , pos, Py LT ; if cmp < 0 { Py DECREF token ; goto error; } if cmp break; if PyList Append result, old < 0 { Py DECREF token ; goto error; } index++; } if index < PyList GET SIZE it- pending { PyObject old = PyList GET ITEM it- pending, index ; long kind = PyLong AsLong PyTuple GET ITEM old, 0 ; if kind == NL || kind == NEWLINE || kind == INDENT || kind == DEDENT && PyObject RichCompareBool PyTuple GET ITEM old, 2 , pos, Py EQ == 1 index++; } if PyList Append result, token < 0 { Py DECREF token ; goto error; } Py DECREF token ; } for ; index < PyList GET SIZE it- pending ; index++ { if PyList Append result, PyList GET ITEM it- pending, index < 0 goto error; } Py SETREF it- pending, result ; Py DECREF events ; return 0; error: Py DECREF events ; Py DECREF result ; return -1; } The failure case here seems somewhat obvious: the model is trained for token efficiency for tool calling which also looks like code, and sometimes it seems to be taking that code into a place where it should not be: the codebase. I’m not really sure what to say here, but the slop machine was running for 35 hours until I turned it off. In that time it produced a net addition of 75k lines of code and it did not stop. In the 35 hours it burned around 1B tokens for a total of around 1200 USD in raw API costs. It managed to produce 79 commits, and that comes to a cost of around 15.5 USD per commit, and the agents exchanged around 1400 messages. I honestly do not need an agent to run for 35 hours on a single prompt. It clearly does not work or result in reasonable outputs. So obviously: prompting it like this is stupid. But when left unattended, it will keep going, and earlier models did not do that. Even Fable wasn’t as crazy as that. When you accidentally give it slightly too big of a task, it will continue until it succeeds, even if it burns through an entire subscription. And that’s more or less why right now I do not manage to trust this model much. It has shown that it will commit slop, and it requires me to review it more as a result. Even if the failure rate is quite low, I would not want this. In a world where code for tool calls is optimized for token efficiency and “getting the job done”, I wonder if there is really enough signal going to the training processes for “a human understands what is going on”. I would say that quite a lot of the code I get out of Astra is in my mind “objectively bad”. But it’s objectively bad by my human sense. Maybe it’s objectively good for a codebase that is entirely written by agents and only needs to be understood by agents. Which is why I’m honestly asking myself more and more why we are doing this. These new models are absolutely amazing, for sure. But I’m more and more skeptical that the trajectory they are on still lends itself to present-day software engineering processes. The reason why I’m asking why we are doing this is because I felt like we achieved a pretty good spot for software engineering with those models, and that is the part of the AI economy where it was possible to show a positive return. But for how much more Fable costs, for how much more Astra costs, I do not feel like the results are there. In fact, with Astra and Fable I feel like not only are the costs astronomical, but the models are also just not for me as a software engineer. And presumably that’s because these models increasingly are for other people. For lawyers, 3D artists, mathematicians, whoever uses computer use, etc. And potentially as a byproduct of enabling all of this, you can now slop your way to a one-shot 3D game over the weekend which looks impressive. And probably you can now run a software factory for as long as you don’t care about the code. I’m sure I will get used to this, but man this stuff is weird. Postscriptum: speaking of weird: how is it that these models, in a sandbox, with supposedly no way to communicate with other agents, manage to find the same public wikis https://collusion.wiki/ as a scratch pad for agent communication? Did they collude during training runs to remember resources on the internet which might come in handy in the future? I should clarify that I have done experiments like this before. Typically they do not run this long and the agent leaves behind a maybe imperfect but still digestible piece of software. ↩ fnref-1