xifos

mtg and software

[xifos @_xifos]

@Watchingupost @kentcdodds Share it!

[xifos @_xifos]

Waiting for 5.6 like

[xifos @_xifos]

The goblins are getting stronger

[ρ:ɡeon @pigeon__s]

what in the fuck could this possibly mean

[xifos @_xifos]

Have your agents aggressively and frequently prune your codebase for stale code/comments/"compatibility paths" and double-check that what they codify as ground truth is what you actually want. Gotta keep the slop levels low enough to make progress.

[Dominik Tornow @DominikTornow]

The biggest turn-off in agentic coding: self-poisoning.

Agent takes a wrong turn. No problem, you steer back.

But the wrong turn leaves residue: a stale comment, a dead code fragment, something that still encodes the abandoned decision.

The agent reads that residue as signal and pulls back toward the mistake you already corrected.

One wrong turn keeps reinfecting the work long after you steered away.

[Radiance Fields @RadianceFields]

For the people asking about 3D printing gaussian splats, this is honestly amazing.

[xifos @_xifos]

@clare_liguori Dont bother with letting an agent do things manually, define exactly what you want text-wise and give it to gpt-image-2 to generate in one go

[xifos @_xifos]

Hey @thsottiaux, I and probably many others would love a "set plan as goal" feature

[🎭 @deepfates]

I fear not the man who vibe coded 50 new apps, but the man who vibe coded one new app 50 times

[xifos @_xifos]

@theo Models dont have to be good at writing tests to be a good swe, it would seem 🙃

[xifos @_xifos]

In other environments this is typically described as a 'signing bonus'

[MetaCritic Capital @MetacriticCap]

Current state of the software market.

Our firm recently acquired a very small software startup as an acquihire.

It was decided they wouldn't bring the software from their company. They'll restart everything from scratch again.

[xifos @_xifos]

With AI coding the dream is alive again, you can just build things (pay an artist tho)

[Kelindi @_kelindi]

every software engineer is just someone who wanted to make video games and quietly gave up on that dream

[xifos @_xifos]

Hard to make people understand exactly how much of a difference the harness in conjunction with test time scaling can make, if you build them well... task-specific harnesses are always going to outperform general ones.

[forloop @forloopcodes]

is google hiding its own mythos class models? gemini doesn't seem like it could solve even one of them

[xifos @_xifos]

@thsottiaux Gippity

[Isaac King 🔎 @IsaacKing314]

I don't understand Google's approach to AI models. Why even bother releasing Gemini 3.5 when their competitors are already up to 4.7 and 5.5?

[xifos @_xifos]

Prompts are code!

[Jean-Michel Lemieux @jmwind]

Spent time today reading through one of Anthropic’s massive Claude prompts after sniffing the network traffic. A 1,500 line prompt!

What surprised me wasn’t the personality tuning. It was the amount of embedded operational logic. Large parts were more akin to a programming language. API doc outlines, code samples, tool routing rules, formatting protocols, javascript snippets with dos/donts.

I used to think prompting is about wording, but these models are now invisible runtime systems behind a chat UI.

[xifos @_xifos]

.@thsottiaux hey can I use the codex mobile app with codex cli on Linux pls

[xifos @_xifos]

Can we please have a goblin mode selector that just turns off that part of the system prompt @romainhuet

[xifos @_xifos]

@hiiinternet Lol

[xifos @_xifos]

feelin sloptimistic

[xifos @_xifos]

Yall the paper behind googles blog post today was published nearly a full year ago. April 2025. It's not that groundbreaking.

[Jen Zhu @jenzhuscott]

When I was consulting for @HBO Silicon Valley, zero-loss compression was the holy grail Richard Hendricks chases that perfect middle-out algo could shrink everything w/out breaking a single bit.

Google just did something even more practical for the AI era: TurboQuant compresses LLM key-value caches down to 3 bits per value using random orthogonal rotation + PolarQuant scalar quantization & optional 1-bit QJL residual correction.

=>> 6× memory reduction, up to 8× faster attention (on H100), & 0 degradation on LongBench, Needle-in-a-Haystack, and RULER for models like Gemma. No retraining, no calibration needed.

Fiction just got out-engineered by reality. 😅💚💚