Babel

[Babel @babelDefi]

@ImNotTheWolf @chetaslua Are you retarded? This is a very well-established testing framework for agents and has been for many years.

[Babel @babelDefi]

@nzmrldev @elliotarledge Wow, look at this

[🍓🍓🍓 @iruletheworldmo]

what if i told you…

bye bye sam.

(sorry for going early but you’ll never figure out who i am so)

[Babel @babelDefi]

@StrawHatCabnBoy @kurtsteiss @fcbtommy123 It’s true I live in Arizona and have been mindblown seeing what he’s been posting

[Babel @babelDefi]

@shade_engine @steph_palazzolo luckily that’s exactly how their business works. They’ll give it to us soon and people will pay them for it.

[Babel @babelDefi]

@tommy5dollar @das_rdsm @kr0der Yeah very true lol

[Babel @babelDefi]

@tommy5dollar @das_rdsm @kr0der I also think the cheaper model would struggle with *taste* in front end work anyway

[Babel @babelDefi]

I work at a much smaller company and have been rewriting our entire company from scratch basically, 10+ app ecosystem and trying to keep the whole system in memory and verifying (lots of screenshots). And I rarely trust sonnet for subtasks etc. I guess maybe I need to start trying out the cheaper models?

[Babel @babelDefi]

I use up my 200/ month limit every single week, usually with a day or two left so I have to switch to my codex for those few days. Running round the clock on 1/4th the usage limits doesn’t sound realistic. I could optimize my usage a bit definitely. But you’re talking 25% of my limit being more than enough

[Babel @babelDefi]

@shade_engine @steph_palazzolo Yeah the model that found 10,000 zero day exploits in nation-critical infrastructure within weeks of being released is probably ass

[Babel @babelDefi]

@nzmrldev @elliotarledge See here

[Taelin @VictorTaelin]

Just an example on Opus 4.8 outclassing GPT 5.5 in ways that are invisible to benchmarks. (When I post these, this is NOT an attack on OpenAI in any way, quite the opposite, I just want things to improve...)

Left both working in a generic goal: "optimize this file".

After 8 hours:
→ Opus 4.8 landed a solid +17%
→ GPT 5.5 landed +30%

I then checked file sizes.
→ GPT 5.5 *doubled* the file
→ Opus 4.6 grew it by 0.1% (!!!!!)

For most benches, 5.5 would have beat 4.8 here, but clearly Opus did a much better job. GPT produced a short-term win that would stale further progress if I merged it. Opus delivered a no-tradeoff, long-term win. And if I had asked GPT 5.5 to "keep the file size the same", it would just start hacking that, minifying, removing docs, etc. - something Opus 4.8 just doesn't do. Its file is as clear as it was when I set the goal.

---

About this: it is an "HVM5 v2" that is even simpler. Now the whole file is at <14k tokens, and consistently outperforming HVM4 by 5-fold. And this version does not have native constructors, only Unit (`()`), Either (`inl(x)/inr(x)`) and Pair (`tup(x,y)`) as primitives, which is very GPU friendly and means we might actually manage to run SupGen on it!

Opus's overnight progress:

[Babel @babelDefi]

@Gabriel78470020 @elliotarledge Wrong. This benchmark will be the only one that really matters moving forward, as we’ve already heard confirmed from basically every lab

[Babel @babelDefi]

@nzmrldev @elliotarledge 4.8 was basically tuned for ultra code. The layers of critique and verification make this pretty believable

[Babel @babelDefi]

@prelife_ @J_D_Becerra @WNY_Plinker @DCplaysgames @Elvish_Harper Are you fucking retarded? Why come into a Star Wars thread to whine about the fact that it isn’t true history and it’s all made up?

[Babel @babelDefi]

@Not_A_DoctorOk @fiago7 @GlynnErnesto Well I live near an airport. There are airports everywhere in the U.S. having a hard time imagining someone who doesn’t see passenger planes in the air daily.

[Babel @babelDefi]

@Not_A_DoctorOk @fiago7 @GlynnErnesto Yeah seeing a passenger plane fly over (the kind that we see 100 of every day) is just as exciting as b52 stratofortress bombers

[Babel @babelDefi]

@bobloblaw635654 @TravLCox @SportsJWQuacken I don’t think I’ve heard about tithing in a sacrament meeting more than a few times in many years

[Babel @babelDefi]

@haider1 Because it is different architecture lol

[Babel @babelDefi]

@KC_Kreative @japan_nobunaga Bumping into someone and telling them “this is unresolved between us” before leaving

[Babel @babelDefi]

@AnthonyGalli @ericernerstedt @TheCriticalDri2 House is really good honestly but it’s not the same genre as breaking bad. It is comfort / background watching like Seinfeld or something

[Babel @babelDefi]

@LostMyHats @AmericaOnlycast I was raised LDS and went to seminary and studied it all through high school etc. even if I don’t consider myself a member anymore, basically none of what you said is true.