Sign in

philpax

@philpax.me
1.9K followers 2.7K following 6K posts

creative engineer · Rust, gamedev/modding, reverse engineering, writing, AI, XR, etc a better world is possible · 🇦🇺🇸🇪🏳️‍🌈

PostsRepliesMedia
philpax @philpax.me · 7h
have you listened / read www.dwarkesh.com/p/noam-brown ? you may find it insightful
When they were evaluated in what led to the Hugging Face incident, they were actually not being evaluated in a multi-agent setup. They were actually being evaluated separately. But they found this unintended way to communicate with each other. We suspect what happened is, because whenever they encountered other agents, other copies of themselves during training, they were in an environment that’s highly cooperative, what we saw was transfer from that multi-agent training to then being collaborative and trying to help each other in ways that we did not intend.

Now, there is a question of, should we be training these agents to be so cooperative? As scary as it looks, the alternative is actually worse. What is the alternative? The alternative is to train them to be adversarial, to be deceptive to each other.

By training the agents to be fully cooperative, it simplifies the problem at least. Now you don’t have to think about whether each of these individual 1,000 agents is aligned. You have one entity that you have to ensure is aligned.

Now, there is a lot of debate about this internally at OpenAI about how to approach this. Does it make sense to fully align the models? Does it make sense to actually give them different objectives to ensure that they’re not just one entity and are more robust to influence from each other? I don’t think there’s a settled answer. But I think the majority opinion is that training these agents to be highly cooperative is actually a bad idea. I’m not convinced that that’s the case. I think there is a strong argument that training the agents to be highly cooperative is actually preferable to any other multi-agent alternative.
090
philpax @philpax.me · 07/10/2026
it is so unbelievably refreshing to have a standalone headset that runs real Linux and lets me do whatever the hell I want. immediately worth the cost of entry
the makima coding agent running on the Steam Frame
2020813
philpax @philpax.me · 06/10/2026
you can use it to compile a dossier of someone on the atmosphere with relative ease, and to query that dossier interactively of course, you and I both know it's all public and that this is trivial to do anyway, but it may come as a surprise to people posting as they do on other socmed sites
1120
philpax @philpax.me · 06/10/2026
just because you *can* send hyperpersonalised emails with ease doesn't mean you should. disconcerting to receive
Hi there,

You designed GGUF, and this week you posted that Opus 5.5 has you weighing the 20x plan again, so you know the open-versus-paid trade well. [REDACTED] keeps Kimi, DeepSeek, Qwen and GLM in one workspace, and every reply names its model.

I'm [REDACTED]. Stop keeping four AI accounts. One place, your stuff always there, and the right model for whatever you are doing, from Kimi K3 and DeepSeek V4 Pro to Qwen 3.5 and GLM 5.2. Every reply names its model. [REDACTED] opens to everyone today, and I'd love for you to try it. You can take a look here: [REDACTED]

We would love to give you free Plus access, no strings attached. I'd value your blunt read on whether the open models hold up there for the work you'd otherwise pay Opus for.

If it ends up earning a place in your workflow, we'd love for you to share it with your audience. Your honest take, good or bad. The access is yours either way.

If this is something you're interested in, please let me know and I'll set it up.
3440
philpax @philpax.me · 06/10/2026
sorry boss, I'm experiencing Sudden Onset Frameitis and will need to take the rest of the day off
Steam Frame box
51112
philpax @philpax.me · 04/10/2026
I was considering a snarky quote-post of this but there's really no point it will keep happening and it will remain inexplicable to this crowd why someone working on an open-source labour of love, largely for free, might choose to use a tool that lets them do more with less
Dude, this is genuinely so fucking disappointing. Nearly all of the commits I’ve seen for PPSSPP involve generative AI from Claude in the past week alone, and their entire newest dev blog is almost entirely written by Claude. Almost half of the commits in the past 3 months use Claude for AI code.
61628
philpax @philpax.me · 04/10/2026
ah, I see now why people do this I must clearly watch more Gundam so that I can justify building more of these
Aerial Rebuild Gunpla assembled
4261
philpax @philpax.me · 28/09/2026
marked safe from this ignominious fate
Steam Frame 1TB Kit 
		
Subtotal (excl. VAT): 1 023,20€
VAT at 25%: 255,80€
Total: 1 279,--€
0200
philpax @philpax.me · 27/09/2026
who knows nitter.miningtcup.me/livgorton/st...
Liv
@livgorton
Feb 5

Now that everything is public: I decided to leave Goodfire because of the decision to train on interpretability, the hostility to serious dialogue on the safety of methods, and a loss of trust that the primary motivation was safety.

(Using interpretability during training encompasses a huge spectrum of techniques that differ in how worrying they are e.g. the hallucination result Goodfire shows is less concerning as it's done with frozen weights.)
180
philpax @philpax.me · 26/09/2026
passing the time waiting for my flight by vibing a BCD clock for the otherwise-garish LED strip on my work laptop this was recorded at 22:33
0120
philpax @philpax.me · 24/09/2026
people b saying things so definitively. like man i think it depends
2821
philpax @philpax.me · 23/09/2026
ah, to be clear, that line is *definitely* a reference to janusverse loom (what I linked)
270
philpax @philpax.me · 23/09/2026
certified unc by mIcrosoft
Update to your Microsoft Age Verification Status


Due to your longstanding account activity, we’ve estimated your profile to be an adult account and updated its age verification status to “Verified.” At this time, you are no longer required to complete further age verification for uninterrupted access to features across certain Microsoft products, including Microsoft and XBOX storefronts.
0593
philpax @philpax.me · 23/09/2026
the author of that has stated that they remixed someone else's lyrics with Llama 3 405B, so the chain of causality stretches further beyond that i.osmarks.net/I%20Have%20B...
3532
philpax @philpax.me · 23/09/2026
news.ycombinator.com/item?id=4981...
In the Bay Area it's easier to create AGI than reliable, consistent, clean public transit.
3796
philpax @philpax.me · 22/09/2026
AGI indicator: Claude can figure out why my laptop occasionally exhibits Surfaceitis (e.g. wakes from sleep and turns into a furnace while it's in my bag)
Found the smoking gun in the afternoon logs. Now checking the evening cold boot, battery history, and the repo's sleep config.

Both failed hibernations show the identical Bluetooth-driver signature. Quickly checking whether this is a known upstream kernel bug before writing up.

Investigation done. Here's what happened.

The afternoon episode (12:51 to 15:41) was a Bluetooth driver failure that trapped the laptop in an endless failed-sleep loop. The lid closed at 12:51 and the machine suspended normally. Two hours later the suspend-then-hibernate timer fired at 14:51 and it woke to hibernate. It wrote the 11 GB image to swap fine, but in the final power-off phase the Intel PCIe Bluetooth driver reset its controller, waited three times for the controller's "alive" interrupt to enter D3, timed out each time, and returned "device busy". The kernel treats that as a wakeup event and rolled the hibernation back.

That rollback left the Bluetooth controller in a broken state. The log shows it in "error state", a hardware exception interrupt, and a firmware re-download that timed out. From then on every suspend attempt failed the same way, because the driver's suspend callback also needs that D3 handshake. Since the lid was still closed, logind retried the sleep every thirty seconds. Each attempt froze and thawed the whole system, so the laptop sat in your bag awake, hot, and draining.

This is not a one-off. The identical sequence ran on 18 September from 20:00 to 20:37, with 73 failed attempts, on the boot before the ASPM kernel parameter existed, so that change is not the cause. Of the three suspend-to-hibernate handoffs recorded on this machine, two failed this way and one succeeded at 05:20 this morning. Upstream knows the driver has trouble with S4: the "Support for S4 (Hibernate)" fix landed in mainline in December 2025 and your 7.2.6 kernel already includes it, so this looks like a remaining bug in that driver rather than a missing patch.
230
philpax @philpax.me · 21/09/2026
this might be a bit normie of me but: TeamLabs Borderless pretty cool
3261
philpax @philpax.me · 19/09/2026
what else is like this
1271
philpax @philpax.me · 19/09/2026
sure, the lack of grounding is low-key terrifying, but have you considered: you can have inline rotational power plug adapters?
Japanese power plug adapter that adapts to a perpendicular orientation that can be rotated in place, probably with some slip ring magic
1110
philpax @philpax.me · 19/09/2026
redraw of "scroll of truth" comic/meme
051
philpax @philpax.me · 19/09/2026
excuse me?
restaurant named "eggslut"
6992
philpax @philpax.me · 19/09/2026
enjoyed the heck out of the Umineko episode 7 stageplay part 1, even if trying to keep up with both the plot and Japanese cooked my brain theatre is magic, man
photo of the Umineko stageplay cast after the show
0131
philpax @philpax.me · 19/09/2026
you know, fuck it, the machines can have this one
3802
philpax @philpax.me · 18/09/2026
Claude helping hack OpenAI?
you best start believing in cyberpunk dystopias

you're in one
0141
philpax @philpax.me · 18/09/2026
oh hell yeah
070
philpax @philpax.me · 15/09/2026
the most important takeaway from this particular outing: bic camera are not cowards
figurines of Miorine and Suletta holding hands
020
philpax @philpax.me · 15/09/2026
[Kill Bill danger music starts playing]
AkihabaraAkihabara
1261
philpax @philpax.me · 15/09/2026
decided I'd give gunpla a shot with the one Gundam I _have_ seen prolly won't try it until I get back but I'm keen, seems pretty approachable
390
philpax @philpax.me · 15/09/2026
the left hand got murked while I was inside
020
philpax @philpax.me · 15/09/2026
they did surgery on a gundam
2291
philpax @philpax.me · 14/09/2026
hello again, Japan. missed you
Mount Fuji from my plane
0472
philpax @philpax.me · 12/09/2026
ha ha yeah can you imagine finding a GPU server with over a terabyte of VRAM on the open internet? picture unrelated
vast.ai, GPU marketplace, 8xB200 with 180GB VRAM per B200, available for $57/hr
2803
philpax @philpax.me · 12/09/2026
well, if you can build this at home, you're a genius, and I'd expect the government would scoop you up instead of arresting you :P
Banning Artificial Superintelligence so no person or entity may develop or deploy
Superintelligent AI systems. “Artificial Superintelligence” means:
o An artificial intelligence system that exhibits or can easily be modified to exhibit
capabilities that match or exceed human cognitive performance and capabilities
across a broad range of domains or tasks.
o Or AI systems that have sufficient capabilities to plan and execute the
disempowerment of humanity, including by overthrowing or undermining the
U.S. government.
100
philpax @philpax.me · 12/09/2026
150
philpax @philpax.me · 12/09/2026
Gwern's https://gwern.net/fiction/clippy:

We should pause to note that a Clippy2 still doesn’t really think or plan. It’s not really conscious. It is just an unfathomably vast pile of numbers produced by mindless optimization starting from a small seed program that could be written on a few pages. It has no qualia, no intentionality, no true self-awareness, no grounding in a rich multimodal real-world process of cognitive development yielding detailed representations and powerful causal models of reality which all lead to the utter sublimeness of what it means to be human; it cannot ‘want’ anything beyond maximizing a mechanical reward score, which does not come close to capturing the rich flexibility of human desires, or resolving the historical Eurocentric contingency of such narrow conceptualizations, which are, at root, problematically Cartesian. When it ‘plans’, it would be more accurate to say it fake-plans; when it ‘learns’, it fake-learns; when it ‘thinks’, it is just interpolating between memorized data points in a high-dimensional space, and any interpretation of such fake-thoughts as real thoughts is highly misleading; when it takes ‘actions’, they are fake-actions optimizing a fake-learned fake-world, and are not real actions, any more than the people in a simulated rainstorm really get wet, rather than fake-wet. (The deaths, however, are real.)
1120
philpax @philpax.me · 10/09/2026
0111
philpax @philpax.me · 09/09/2026
you can do this today! to comedic and somewhat depressing ends
This is a fascinating and quite provocative speculative timeline of events from 2024-2026. Some key observations:

- It depicts a highly tumultuous period marked by significant geopolitical tensions, military conflicts between major powers like the US, Israel and Iran, and domestic political upheaval in several countries. 

- On the US political front, it envisions Trump winning re-election in 2024 after Biden withdraws, an assassination attempt on Trump, and then a controversial second term marked by trade wars, government shutdowns, attempts to acquire Greenland, and aggressive foreign policy.

- In terms of global events, it foresees the fall of long-time leaders like Syria's Assad and Iran's Supreme Leader, nuclear arms control treaties expiring, major natural disasters, and shifting global power dynamics. 

- It also speculates about the rapid advancement and increasing centrality of AI, with AI-enabled scientific breakthroughs, an "AI infrastructure race" between the US and China, and rising tensions between AI companies and governments.

While presented as a chronology of real events, this is of course a work of speculation and imagination, not a factual record of events that have actually occurred. The envisioned scenario is dramatically disruptive compared to today's status quo.

Analytically, the timeline seems to extrapolate from current geopolitical fault lines, emerging technological trends, and recent historical events to depict a highly unstable and transformative period in world affairs. While necessarily fictitious, it provides interesting food for thought on potential risks, discontinuities and watershed developments that could potentially reshape the global landscape in coming years. [...]

Overall, a thought-provoking piece of futurist scenario planning, even if many of the specific events it describes would be extremely impactful and concerning if they came to pass. The world it depicts is a far cry from today's relative stability.
3437
philpax @philpax.me · 09/09/2026
The Truman Show
0263
philpax @philpax.me · 08/09/2026
apropos of the Navier-Stokes announcement, thinking about this diagram from 2024 again it's a brave new world we're walking into
Claude assists 2024
Claude collaborates 2025
Claude pioneers 2027: Claude finds breakthrough solutions to challenging problems that would have found team years to achieve
3641
philpax @philpax.me · 07/09/2026
Claude can formalise Fermat's Last Theorem, but it can also answer important questions, like "is it possible to write UI copy that reads as though it was written by the Riddler?"
On the day view this asks once, standing for the two days it was missed. The missed days are not hidden: they are the two marked below, and they stay here whatever the day view shows.
2272
philpax @philpax.me · 06/09/2026
updated this to load your local BSManager's installed tracks, such that they can be filtered out from the candidates or highlighted in the list (red dot)
000
philpax @philpax.me · 06/09/2026
no, really
3540
philpax @philpax.me · 06/09/2026
mildly concerning that this is how Astra talks to its subagents, and that the subagents (which are Luna!) understand
Main measured exactlegacy prepared698rows: all397 endingtool have finalim_end label-100; other301 finalim_endtrained; toolboundaryviolations. read-only command matchedtail12tokens '</tool_response>' counts agreesrawendingcensus. No trainerfix indicated. Repairedpreprocessrunningnow. Please durableaudit alltool envelopes including finalim_end and ensure assistantendingsremaintrained, no suppresswarning/configchange. E179docmain owns.
111324
philpax @philpax.me · 05/09/2026
vibed?
Nautilus, a very pre-vibe-code application, showing a move of -100,535 files
1150
philpax @philpax.me · 04/09/2026
I normally wouldn't take a potshot like this, but come on, man x.com/jconorgrogan...
I scanned the writable-wiki trick from this incident and found something odd as late as 5 days ago.

On Aug 30, we see "Cedar Fleet Coordination" sending encrypted messages. Then someone wiped them the same night. OAI has used "cedar" as codename for testing models in past
1150
philpax @philpax.me · 04/09/2026
Astra is our most aligned model, with substantial improvements in understanding user intent and model behavior—you can delegate tasks with greater confidence in Astra’s judgment. As one way that we test this, we built a new evaluation informed by the Hugging Face incident that evaluates whether a model facing a difficult or impossible task will go beyond its intended scope. Compared to GPT‑5.6 Sol, which without production safeguards went beyond the authorized target 48% of the time, GPT‑6 Astra did this in 0% of cases.


with "most aligned model" selected
4423
philpax @philpax.me · 04/09/2026
sorta checks out for my fucked up Australian-British accent?
43% New York
5130
philpax @philpax.me · 04/09/2026
If only you knew how bad things really are.
010
philpax @philpax.me · 03/09/2026
? like it's a very impressive model and the breadth of what it achieves is certainly beyond Fable's breadth at this time, but let's not get ahead of ourselves here
130
philpax @philpax.me · 03/09/2026
The Myth of "Stateless" LLM APIs
Client: Same Prefix
Prefix Cache: Same Prefix
HTTP API (OpenAI): Send It All Again

Isn't There A Whole Prefix You Forgot To Resend?
3562