xyzzy_plugh

I actually laughed out loud.

Ok so they're a fairly large organization. Compared to other similarly sized tech companies, they pay pretty far below market rates. They're not exactly looting the coffers.

They are doing pretty well financially, yes, and they have done well with their endowment, but it's not like they have an order of magnitude more money then they are letting on.

Are their banners obnoxious? Yes. Are they effective? Apparently.

What would you rather have? Do you really want Wikipedia to be destitute? "Hey guys Jimmy here we don't really need any money but I guess you can give us some?"

No one knows what the future holds. A strong endowment fund today isn't a guarantee for tomorrow.

I'm glad they're striking when the iron is hot. I hope they build up enough wealth to operate for generations upon generations. Of the things on the modern Internet that offend my senses, the fundraising banners on Wikipedia probably rank dead last.

I don't get how people can be so negative about such an overwhelmingly positive contributor to society.

edit: you know what? Fuck it: I hope they make far more than enough money to pay everyone who works there an incredibly comfortable salary so they can all focus in improving humanity's knowledge. You just convinced me to donate.

I remember seeing Fury Road in theatres with tempered expectations. It was a phenomenal experience. We were very surprised at how good of a movie it was.

Mad Max is big shoes to fill, and Fury Road managed it in a way that was so fresh I barely thought about the previous films. It truly stands alone and that's no easy feat. I do not find it derivative at all. If anything it pulled the franchise in a new direction.

Everything is just awesome. They made an awesome movie. The cars are awesome. The characters are awesome. The sounds and practical effects are awesome. The pace is awesome. It's just great stuff.

It surpasses mediocrity across all metrics. That's why it's more than "fine".

"It all just goes to shareholders"

What is this nonsense? It's a non-profit. There are literally no shareholders.

Just send an email when you post a new blog post.

Please create a newsletter sign up or something. I have several clients who could be interested in this as it matures.

The problem is that when you build at this layer, you end up writing far more scheduling code than business logic. Eventually you've built the abstractions we already use today.

The vast, vast majority of software, and software developers, simply do not care about core affinity, NUMA, noisy neighbors, etc.

Even on systems or situations where this matters (massive number of cores and memory, multiple sockets, thousands of threads, etc.) the juice is rarely worth the squeeze.

Doubly so when you end up being I/O bound anyways.

Nearly every software developers wants their software to work across multiple deployment targets. Maybe you just target x86-64, and you don't care about the number of CPUs because you'll be upgrading your servers every couple of years.

No one wants to deal with difficult-to-reproduce issues related to scheduling in niche configurations. It's much more efficient for runtime or compiler engineers to solve these problems for us, leveraging the vast amount of resources of corporations to develop and test these features. Then you have a place your code can live happily ignorant of its runtime constraints.

And that, my friend, is the beauty of our industry.

They're not very accurate. They are generally accurate with a margin of error big enough that they're far from a sure thing. If they were accurate then no one would take a contrarian position. I'd look at a prediction and say "well obviously the favored side is going to win" and put all my money there, over and over again, until I was the richest mortal in the universe.

Polymarket has it at ~50% that their IPO value is between 1.25T and 1.5T. Do you think that is representative of reality? It's not like we get to flip the IPO coin ten times.

I've mentored a lot of young engineers and something I often point out is that there is practically zero risk in accepting an offer. Many companies also do not even present the full paperwork to sign until your first day or two.

You don't have to show up on the start date, though you do have to be mindful of the bridge you may be burning. But it is totally reasonable to show up to a few days of work, and declare "you know what, I don't think this is a good fit" and no one will be offended. They don't have to know that you already accepted another offer.

Exploding offers, low-ball offers, all bullshit. But you don't have to play their games.

Leveraging compromised devices is a very important part of the Internet??? This is an incredibly twisted perspective.

Help me understand. As far as I'm aware, these residential proxies are almost always malware.

Yes and that would make sense, and what I would expect, if it were available. I understand it is load-bearing but it is nonetheless surprising that it hasn't been revisited. It is surprising, at least to me, that time-travel never shipped given its feasibility.

Yes. This is the thing that tripped me the most when learning postgres internals. It was obvious that append-only enabled stable addressing but then it just never materializes. I was incredibly perplexed. Surely I'm missing something right? Nope!

A long time ago I worked with Aphyr to evaluate a database with Jepsen. He was extremely professional, and he found a few real issues which we resolved extremely urgently. We took things seriously and wrote a significant amount of testing ourselves, and by the time our engagement began, we were pretty confident. Overall it was a great experience and I highly recommend his services.

By comparison, this report paints an absolute clown show. The amount of anomalies is perhaps understandable (a single architectural flaw can have a lot of impact) but the fact that they found so many disparate issues is telling.

The fact that they then said they addressed the issues, but didn't, is about the biggest red flag imaginable. I would never let any client use this software under any circumstances.

It really says something that a 404 can get to the front page in 20 minutes.

Raw license text:

https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/raw...

Very much not open as in open source. It's a shame an organization such as OSI doesn't get out ahead of licensing of weights, what do we call this? "Weights Available" ?

This is pretty funny. As someone who has overseen the purchase of thousands of high end LGs and maybe a hundred Apple monitors, we saw many, many more issues with the Apple monitors.

Their lower tier monitors have issues but so do literally every other manufacturer. The high-end LGs are very good and QC as of 2026 has been phenomenal.

I tend to agree with this perspective however I'd like to point out that using token quantity and cost as a measure of productivity is counter-intuitive. The variance for a given task can be extremely high regardless of the input, not to say that there isn't any skill to prompting or that talent doesn't matter.

Ultimately it feels pretty similar to using lines of code as a measure.

Instead I recommend setting a budget per head and evaluating performance traditionally, allowing for exceptions in exceptional circumstances of course.

There's a lot of nuance in reverse proxies. You're off to a good start with Go as the standard library gives you a very strong and battle-tested HTTP implementation as well as a basic reverse-proxy interface to build with.

Avoiding opening a can of security-issue-worms involves knowing what you're doing and depends heavily on the environment you're operating in. Internally proxying traffic between services? A cakewalk. Translating between HTTP2 or 3 to 1.1? Forwarding origin client IP safely? Sanitizing headers (and I don't mean the in-built sanitization that Go provides)? There's a lot of little things.

Tools like Nginx and Envoy strike from the opposite direction: an extensive suite of configurable tools to reverse proxy the way you need, with relative safety.

Like you, I would almost certainly prefer writing a small reverse proxy in Go over deploying Nginx.

Filippo, I'm rather surprised and disappointed to see this come from you. I'd much rather see Google working on preventing abuse of apps already installable via the Play Store, instead of imposing restrictions on those who do not want to use their store to begin with.

I don't mind warnings and even brief cool downs but 24 hours is ridiculous.

If I lose or break my phone, I have to wait 24 hours after acquiring my new phone to use the apps I previously used? It's just absurd.

If I had to explain to clients that I was unable to do work until tomorrow, I'd be ridiculed.

I'm gonna be real with you. No one needs to know docker syntax. Most engineers just copy and paste chunks of Dockerfiles around anyways.

Docker syntax is one of those things that just immediately falls out of my head when I close the editor. I don't care about it at all.

Even better: there are many, many tools that build OCI images without using the Docker syntax. I haven't hand-written a Dockerfile in at least six years.

LLMs have not loved the needle here in my sphere, it was already bouncing off of the zero on the gauge.

edit: this isn't to say the syntax is good or bad (though there are several questionable parts of it) but extremely simple, and concise enough that you can just look up literally any command with great ease. It's surely a large part of its success, for better or for worse.

So it's like Archil or ZeroFS but more proprietary and less useful. Got it.

Look, we get it: anyone can write a filesystem now. It doesn't mean you should.

I don't believe you.

Here, I inserted a subtle bug in this function, can you find it?

    int widget_grid(lv*target,grid x,grid_val*value){
        if(x.show==show_none)return 0; lv*hfnt=FONT_BODY; int hsize=x.headers?font_h(hfnt)+5:0; if(x.size.h<=(50+hsize)||x.size.w<16)x.scrollbar=0;
        lv*fnt=x.font?x.font:FONT_MONO; int os=value->scroll, or=value->row, oc=value->col, files=x.headers==2; if(files||x.size.h<=hsize)x.headers=0;
        int _bg=dgeti(value->table,lmistr("_bg")),_fg=dgeti(value->table,lmistr("_fg")),_hideby=dgeti(value->table,lmistr("_hideby"));
        lv*pv=NULL,*vp=NULL;if(target!=NULL&&_hideby!=-1)grid_pv(target,&pv,&vp);int nr=pv?pv->c: value->table->n;
        #define grid_cell_at(col,row) (value->table->lv[col]->lv[permuted_row(row)])

        int nc=MIN(256,value->table->c), rh=font_h(fnt)+(x.lines?5:3), fcol=!in_layer()?13:x.show==show_invert?32:1, bcol=x.show==show_invert?1:32;
        int fk=strlen(x.format); rect b=x.size;
        int sel=in_layer()&&x.show!=show_none&&wid.active==wid.count;
        if(in_layer()&&dover(b)&&(ev.md||ev.mu||ev.drag)){if(!sel&&wid.gv)grid_exit(); wid.active=wid.count,sel=1;}
        if(sel&&in_layer()&&!over(b)&&ev.md){sel=0;wid.active=-1;grid_exit();}
        if(sel){if(wid.fv)field_exit();wid.ingrid=1,wid.g=x,wid.gv=value,wid.gt=target;}
        if(x.show!=show_transparent)draw_rect(b,bcol);
        rect bh={b.x,b.y,b.w,x.headers?font_h(hfnt)+5:0}; int nrd=MIN(nr,((b.h-bh.h+1)/rh)), scrollmax=nr-nrd;
        rect bb=scrollbar((rect){b.x,b.y+(x.headers?bh.h-1:0),b.w,b.h-(x.headers?bh.h-1:0)},scrollmax,1,nrd,&value->scroll,x.scrollbar,x.show==show_invert);
        draw_box(x.lines?b:(rect){b.x,bb.y,b.w,b.h-(bb.y-b.y)},0,sel?13:fcol);
        char*pal=patterns_pal(ifield(deck,"patterns"));
        if(x.lines)draw_rect(bh,fcol);if(nc<=0)draw_textc(inset(bb,1),"(no data)",hfnt,fcol);

        int cw[256];for(int z=0;z<nc;z++)cw[z]=z>=x.widths[0]?-1:x.widths[1+z];
        if(_bg!=-1)cw[_bg]=0;if(_fg!=-1)cw[_fg]=0;if(_hideby!=-1)cw[_hideby]=0;
        int hwt=0;for(int z=0;z<x.widths[0];z++)hwt+=cw[z];
        int nwt=0;for(int z=0;z<nc         ;z++)nwt+=cw[z]==-1;
        for(int z=x.widths[0];z<nc         ;z++)cw[z]=cw[z]==-1?(bb.w-hwt)/nwt:0;

        if(!cf.sv)cf=str_new();
        #define rowh(n) inset((rect){bb.x+1,bb.y+rh*n+2,bb.w-2,rh-3},x.lines?0:-1)
        #define rowb(n) (rect){bb.x,bb.y+rh*n,bb.w,rh}
        int clicked=0,rsel=0,hrow=-1,hcol=-1;
        for(int y=0;y<nrd;y++){
                if(_bg!=-1){
                        int p=CLAMP(0,ln(grid_cell_at(_bg,y+value->scroll)),255);
                        if(p){rect t=rowb(y);if(y==0)t.y+=1,t.h-=1; draw_rect(t,p);}
                }
                int ra=in_layer()&&over(bb)&&over(rowb(y));rect cbox={0};
                if(ra&&x.bycell){for(int z=0;z<nc;z++){rect cell=grid_cell(x,value,bb,(pair){z,y},cw);if(over(cell))hcol=z,cbox=grid_hcell(x,z,cell);}}
                if(ra&&(ev.md||ev.drag)){rsel=1,hrow=y+value->scroll;draw_rect(x.bycell?cbox:rowh(y),fcol);}
                if(ra&&ev.mu)clicked=1,value->row=permuted_row(y+value->scroll),value->col=hcol;
                if(ra&&!ev.drag)uicursor=cursor_point;
        }int rr=display_row(value->row)-value->scroll;
        if(!rsel&&rr>=0&&rr<nrd&&(x.bycell?value->col>=0&&value->col<nc:1)){
                hrow=display_row(value->row),hcol=value->col;
                if(x.bycell&&value->col>=0){draw_rect(grid_hcell(x,value->col,grid_cell(x,value,bb,(pair){value->col,rr},cw)),fcol);}
                else{draw_rect(rowh(rr),fcol);}
        }
        int drawncol=0;for(int z=0,cols=0,cx=0;z<nc&&cx+cw[cols]<=bb.w;z++,cols++){
                rect hs=(rect){bh.x+4+cx,bh.y+1,cw[cols]-5,bh.h-2};
                if(hs.w<=0)continue; // suppressed column
                if(x.headers){
                        int oa=target&&in_layer()&&over(hs)&&((ev.drag||ev.mu)?dover(hs):1)&&!wid.col_drag;
                        int dp=oa&&(ev.md||ev.drag);if(dp)draw_rect(hs,x.lines?bcol:fcol);
                        draw_textc(hs,value->table->kv[z]->sv,hfnt,x.lines^dp?bcol:fcol);
                        if(oa&&!ev.drag)uicursor=cursor_point;
                        if(oa&&ev.mu){msg.target_order=target,msg.arg_order=value->table->kv[z];}
                }
                if(drawncol&&x.lines)draw_invert(pal,(rect){hs.x-3,b.y+1,1,b.h-2});cx+=cw[cols];drawncol=1;
                for(int y=0;y<nrd;y++){
                        int ccol=y+value->scroll==hrow&&(x.bycell?cols==hcol :1)?bcol: _fg==-1?fcol: CLAMP(0,ln(grid_cell_at(_fg,y+value->scroll)),255);
                        rect cell={hs.x-3,bb.y+rh*y+1,hs.w+5,rh-1}; lv*v=grid_cell_at(z,y+value->scroll);
                        cf.c=0;format_type_simple(&cf,v,z>fk?'s':x.format[z]=='L'?'s':x.format[z]);str_term(&cf);
                        rect ib=box_center(cell,image_size(ICONS[0]));pair ip={ib.x,ib.y};
                        rect oc=frame.clip; frame.clip=box_intersect(cell,frame.clip);
                        if     (z<fk&&x.format[z]=='I'){int i=MAX(0,MIN(8,ln(v)));if(i<8)draw_icon(ip,ICONS[i],ccol);}
                        else if(z<fk&&x.format[z]=='B'){if(lb(v))draw_icon(ip,ICONS[icon_chek],ccol);}
                        else if(z<fk&&(x.format[z]=='t'||x.format[z]=='T')){
                                rect rr={hs.x+1,bb.y+rh*y,hs.w-2,rh}; // display rich text:
                                pair t=layout_richtext(deck,rtext_cast(v),fnt,align_left,rr.w);
                                rect r={rr.x,rr.y+ceil(t.y<rr.h?(rr.h-t.y)/2.0:0),rr.w,rr.h};
                                lv*alink=NULL;if(x.format[z]=='T'&&in_layer()&&(ev.md||ev.drag||ev.mu))for(int z=0;z<layout_count;z++){
                                        glyph_box g=layout[z];if(g.pos.w<1)continue; // skip squashed spaces/newlines
                                        if(g.pos.y+g.pos.h<0||g.pos.y>r.h)continue; g.pos.x+=r.x, g.pos.y+=r.y; // coarse clip
                                        if(lis(g.arg)&&g.arg->c&&dover(g.pos)&&over(g.pos)){
                                                if(ev.md||ev.drag){alink=g.arg;break;}
                                                if(ev.mu){msg.target_link=target,msg.arg_link=g.arg;break;}
                                        }
                                }
                                draw_text_rich_raw(r,ccol,0,alink);
                        }
                        else{
                                rect r={hs.x+1,bb.y+rh*y,hs.w-2,rh}; // right-align numeric columns:
                                if(z<fk&&strchr("fcCihH",x.format[z])){draw_textr(r,cf.sv,fnt,ccol);}else{draw_text_fit(r,cf.sv,fnt,ccol);}
                        }
                        frame.clip=oc;
                        if(!x.locked&&sel&& ((ev.dclick&&over(cell)) || (ev.action&&x.bycell&&z==value->col&&y+value->scroll==display_row(value->row)))){
                                char f=z<fk?x.format[z]:'s';pair tc=(pair){z,y+value->scroll},tpc=(pair){tc.x,permuted_row(tc.y)};
                                if     (f=='I'||f=='L'||f=='T'){} // no editing allowed
                                else if(f=='B'||f=='b'){grid_edit_cell(tpc,ls(lmn(!lb(v))));} // toggle
                                else{
                                        ms.pending_grid_cell_rich=f=='t';
                                        wid.pending_grid_edit=1,ms.pending_grid_cell=grid_cell(x,value,bb,(pair){tc.x,tc.y-value->scroll},cw);
                                        ms.pending_grid_cell.y-=1,ms.pending_grid_cell.w+=1,ms.pending_grid_cell.h+=2;
                                        ms.cell=tpc;ms.text=(field_val){f=='t'?rtext_cast(v):rtext_cast(lmcstr(cf.sv)),0};
                                }
                        }
                }
        }
        if(x.lines)for(int y=1;y<nrd;y++)draw_hline(bb.x,bb.x+bb.w,bb.y+rh*y,fcol);
        if(!x.locked&&in_layer()&&target)for(int z=0,cx=bh.x;z<nc;cx+=cw[z],z++){
                if(cw[z]==0)continue;
                rect h={cx+cw[z]-1,bh.y,5,bh.h};if(h.x+h.w>b.x+b.w)break;
                if(over(h))draw_vline(h.x+2,h.y,h.y+h.h,13);
                if(ev.md&&dover(h)){wid.col_drag=1,wid.col_num=z,wid.col_orig=cw[z];}
                if(sel&&wid.col_drag&&wid.col_num==z&&ev.drag){
                        int s=MIN(MAX(10,wid.col_orig+(ev.pos.x-ev.dpos.x)),bb.w-10),i=z;uicursor=cursor_drag;
                        GEN(r,MAX(x.widths[0],i+1))lmn(i==z?s:cw[z]);iwrite(target,lmistr("widths"),r),mark_dirty();
                }
        }
        if(target&&os!=value->scroll)iwrite(target,lmistr("scroll"),lmn(value->scroll)),mark_dirty();
        if(target&&or!=value->row)iwrite(target,lmistr("row"),lmn(value->row)),mark_dirty();
        if(target&&oc!=value->col)iwrite(target,lmistr("col"),lmn(value->col)),mark_dirty();
        if(target&&clicked)msg.target_click=target,msg.arg_click=(fpair){0,value->row};
        if(in_widgets())wid.count++;
        return files?((clicked&&ev.dclick)||(sel&&ev.action)): clicked;
    }

I can't resist posting whenever Decker shows up.

Decker is one of the craziest code bases I'm aware of: a single, incredibly dense file:

https://raw.githubusercontent.com/JohnEarnest/Decker/refs/he...

(I wonder how many tokens it'd be to make this legible...)

I've seen three outages at various orgs that were prolonged due to compile times. In these cases it was a root contributor to missing RTO. The compile times are a liability.

One proposal I came across actually suggested breaking up a perfectly fine monolith into a dozen microservices to reduce compile times (at the cost of cumulative CI costs going up).

When the alternative is massively rearchitecting your systems, Zig and Go begin to look really nice by comparison.

Yes. The area on the Venn diagram where services, hardware and regulatory compliance intersect is surprisingly small.

I had a client that was contracting with a very large bank and there were only something like four zones globally that we could possibly deploy in to.

Google also is unlike AWS in that their zone abstraction is very virtual, and two zones might be in the same physical building. If you have requirements that your deployment zones be physically separated by some minimal distance then it can make things pretty challenging.

No.

If this is endorsed and thus malicious, they should be ground into dust by the law.

If this is accidental and due to incompetence, they should be ground into dust by the law.

I've secured many, many far less important and impactful systems substantially better than even the best sandbox environment that we're aware of being used by these labs.

There's simply no reason to let these things interact with the internet at large. We can already rapidly copy existing tools and services (albeit questionably) using these models, why not simply run in a fully virtualized, air gapped environment?

Because they don't care.

This is incredibly blatant intentionally malignant behavior. The fact that so many people question this is frankly stunning.

They're getting away with it simply because they can point at the other guy and say "see? they're doing it too!"

It's ridiculous and they should all be litigated into a fine residue.

We take biological lab leaks seriously, how is this substantially different?

I once worked on a team of SREs where the workload was so atrocious that we had HR make up new virtual employees Team On-call Primary and Team On-call Secondary and, other than the manager, they were the only accessible team members. We gave these virtual employees their own laptops and phones. When your shift ended, or you needed a break, you handed the gear to the next guy. We didn't have a formal schedule because it was common to get burned out in just three or four days, or if it was a rare calm period you might hang on to the laptop a bit longer until you got a chance to get properly throttled.

It worked exceptionally well because then the other employees were unable to individually bother team members. The team name in the HR directory was something vague and reported to the skip level, so unless you were in the know, we were totally invisible to the rest of the org. For some context we were keeping many millions of dollars of daily transactions afloat and it was not unusual to receive death threats from business folks when they couldn't do their jobs.

I don't miss it remotely, but I sometimes miss being a member of an invisible force with superuser privileges, with the blessing to unfuck anything we saw fit.

That would mean they would be growing slower. Being common is irrelevant.

Raising a Series D is an attempt to permanently increase the velocity of their enterprise. The change in the cap table is negligible if they succeed.

Raising debt right now would be a massive unforced error given they were able to shore up half a billion dollars in funding. If the funding well were dry then it'd be a different story, but that's not the case.

Can you imagine if they raised $445M of debt?

This is a crazy take. If you can raise on good terms (or better) then it almost always beats debt. Debt can make it a lot harder to raise or take on more debt in the future.

They're profitable, they are going to get the best terms possible at this moment.

I like gvisor but I fear this is too little, too late. Microvms are crushing it right now. I believe even Modal, which is one of the few non-Google users deploying gvisor commercially, and heavily called out in TFA, are moving off gvisor in favor of microvms. Unreal!

Frankly I would never want to operate gvisor with customer defined workloads.

For one, performance isn't native, which makes it hard to determine workload characteristics. Second, it has a pretty long list of limitations that make it difficult to deploy transparently: cgroups aren't fully supported, no io_uring, etc.

They claim they test a wide array of workloads but there are still compatibility issues that take years to iron out. I once had a workload that was probably buggy, and crashed only on arm64 under gvisor, but as I lacked source access there was no viable path forward.

I hope I am wrong, I'd love to see gvisor really take off. I think finishing the macOS port would be wild: run Linux binaries on macOS without a VM! But I'm not confident.

This is what happens when you give people a valuable service for free. You can't take it away without a lot of complaining and loss of good will.

Flash-lite is pretty capable and getting that entirely for free is not a bad deal at all.

It was always obvious Google couldn't keep letting non-paying users use Flash and Pro.

I'm not a fan of how they bundled their plans but they're not too bad especially when shared with a family. I think I've got almost 1TB of Google Photos at this point.

This is a much better page than TFA: https://support.google.com/gemini/answer/17004136