wren6991
wren6991 on Hacker News
It's confusingly worded but I think it's a suggestion for something Apple could do in the future. I don't see any mention of this in the linked Apple article.
Also a great way to make money from screen replacements when the battery expands!
Yes, V4.1 Flash 1017 is my daily driver cloud model. Copes fine with what I'd consider difficult brownfield development, like fixing bugs in vLLM prompt cache and coming up with small/sane overlay patches for them. Maybe it's disappointing for people who want to one-shot a game or website from a single prompt; I haven't tried that.
It's nice that their peak-rate hours end quite early in the day UK time.
They're still on 3.1 while Anthropic are on 5.5 and OpenAI are on 6. Insane how far behind they're getting.
> Harness have network access by default since they are driven by their provider API
The harness needs network access, but the shell spawned by the harness that runs the agent's commands does not necessarily need it.
Having a bwrap etc managed by the harness (DSH does this, among others) lets you have tighter restrictions on the agent's shell than the harness. Then run the whole thing inside a VM with a private network, so you can add nft rules to let it dial out to your inference machine but not to the internet at large. It takes an afternoon to set up but gives some reasonable defence in depth.
How did you go about it?
Fun fact: OpenAI Dot and Claude Duet are both canonically pronounced "doot"
Not that uncommon for a language interpreter? Pretty sure MicroPython does some setjmp/longjmp for non-local returns.
Did an LLM do the write-up? Half of the "technical deep dive" was just re-stating earlier statements about scope and impact, and I left without a solid grasp of the mechanics of the exploits.
Tried it. Little whale is getting smarter. Or, as DeepSeek might say: "Hmm. Wait! Hmm hmm."
As a philosopher^H person on twitter once said: "If you find two ants in your kitchen, the best estimate of the total number of ants in your kitchen is not two." OpenCode has some pedigree here.
I guess for me there is a distinction between an agent I am using interactively and walking through step-by-step, vs an agent that is running unattended, doing web search, communicating with subagents etc. I am ok with giving greater privilege to the former. But yes Qwen-chan accidentally committing suicide, or OOMing its own inference machine is a canon event.
One of the things I hit is that there is no way to trigger an "open folder" dialog. It gets mapped to "open file" instead. The correct hooks exist in both WebKit and in GTK, but not wired up in WebKitGTK. I think the fix would be quite minor but I looked into what is involved in raising a patch and decided to leave it. There are a lot of minor paper cuts like this.
I did hit a couple of CSS or rendering quirks, but the rendering perf also seemed noticeably poor, and that's harder to live with.
Right. The US has famously never installed government-mandated backdoors in security-critical technology like Dual_EC_DRBG.
Using Tauri on Linux gave me appreciation for why Electron just ships an entire browser. WebKitGTK is kinda rough around the edges.
It's a shame the RLVR training doesn't zap the model with a million volts every time it runs a recursive grep or find on /, but I guess that would make the benchmarks suffer.
> you cannot have both and be serious about security
Why not? I think DSH does this pretty well with an agent preset that exposes the tools required for the agent to modify the harness.
It's kind of like how the kernel can expose interfaces to allow userspace to load kernel modules at runtime: you "just" have to control which processes can use them.
Between that, and a couple of extra injected bwrap args to make my SSH keys and LLM API credentials unreadable from the agent's shell by default, I'm pretty comfortable using the harness outside of a VM for most things.
Agreed, DSH is pretty great
I was able to beat this quite easily by just saying the first 100 digits of pi out loud. The house guessed 1/10, no better than random. Maybe you should add a predictor for "well-known mathematical constants".
They're trying to distil OpenAI employees now. Absolutely shameless
When I had a Claude account (cancelled ~May time) I vividly remember the web chat having broken editing for backtick-quoted monospace text. If you typed <space>`foobar`<space> the last space would move the cursor backward into the closing quote, and the space would be emitted inside the monospaced region. It was like this for several months (on Firefox). I always found this baffling for a product marketed to programmers.
If software engineering is solved they should be able to get it right the first time
It just reads like vagueposting about how capable internal models are. We've been hearing this for a while, yet each model that's actually released is a modest increment on the previous. The boy keeps crying wolf.
If you look at this and you think "wow, I love the idea of CAD-as-code but I wish it was a more sane language" then have a look at CadQuery:
https://cadquery.readthedocs.io/en/latest/quickstart.html#qu...(I recommend CQ-editor for an easy install: https://github.com/CadQuery/CQ-editor/releases/)
It's a similar idea to OpenSCAD, but embedded in Python and using a better geometry kernel. Want to fillet a box in OpenSCAD? Sure, either take the convex hull of 8 spheres, or the minkowski sum of a box and a sphere. Want to fillet a box in CadQuery? .edges().fillet()
Now I'm hearing some protests about Python, but consider this example OpenSCAD code:
x = 1;
echo(x);
x = 2;
In most languages you would expect this to print "1". In OpenSCAD,
the last assignment wins and it prints "2".
Read the tech report and see how laser-focused they've been on compressing the disk KV footprint specifically. 890 bytes/token is absurd, and is probably the reason I regularly get ~0.5 Mtok request full cache hits after an hour. On their end that's just yanking a 414 MiB file off an SSD, then doing a little bit of compute (bounded SWA replay). Nobody else seems to be serving their model as well as they do.
Yes, I send it to China, and they use it to improve open-weight models. I'm ok with this arrangement. At least, I'm happier with this than with companies using my open-source work to improve proprietary models without my consent.
It's a solid little model, and I appreciate DeepSeek's commitment to the bit in releasing a brand new pretrain, double the size, numerous architectural innovations as a ".1" release over the excellent DeepSeek V4 Flash.
Are you counting the n-gram/PLE as part of the model weights there? They can go in host memory. Would be good to show your working. Also the released weights are pre-quantised and presumably QATed, so your "Full Precision" and INT8 are simply not a version of the model that actually exists.
Edit: I went and checked for you. The LM backbone is 307.2 GB (286.1 GiB), straight from DeepSeek's upload. The n-gram table is 203.1 GB (189.1 GiB), which goes in host RAM. Note the embeddings are higher precision than the expert tensors, so it's a larger fraction of the bytes than it is of the parameters.
So,
> Call me crazy but:
You're crazy. :-)
If WINE is not an emulator then YAML ain't a markup language
> There's kind of a bitter lesson variant with AI coding
Yeah, it feels like we've entered the era of disposable software.
Also I do like how everyone in this thread seemed to immediately and unanimously clock this project as slop.
Yeah, maybe that's the difference. I was trying to probe the semantics, not so much a "gotcha". Parts of this stack are quite bespoke for this one game.