conputer dipshit
LightNews LightNews
banner
davidcrespo.bsky.social
conputer dipshit
@davidcrespo.bsky.social
1.5K followers 520 following 4.7K posts
web dev + hot dad. enjoy charts, unions, conputer games, philosophy. chicago crespo.business
Posts Replies Media Videos
davidcrespo.bsky.social
arnold from the magic school bus
November 26, 2025 at 3:51 AM
davidcrespo.bsky.social
$ ai give me a cornbread recipe. no bs, just ingredients and instructions gemini-3-pro | 7s | $0.00581 | Tokens: 221 -> 447 ### Ingredients • 1 cup yellow cornmeal • 1 cup all-purpose flour • 1/4 cup granulated sugar (adjust to taste) • 1 tablespoon baking powder • 1 teaspoon salt • 1 cup milk or buttermilk • 1 large egg • 1/3 cup vegetable oil or melted butter (plus extra for greasing) ### Instructions 1. Preheat: Set oven to 400°F (200°C). If using a 9-inch cast-iron skillet, place it in the oven while it preheats. Otherwise, grease a 9-inch square baking pan. 2. Mix Dry: In a large bowl, whisk together the cornmeal, flour, sugar, baking powder, and salt. 3. Mix Wet: In a separate bowl, whisk the egg, milk, and oil (or melted butter) until combined. 4. Combine: Pour the wet ingredients into the dry ingredients. Stir just until moistened; do not overmix (a few lumps are fine). 5. Bake: Carefully remove the hot skillet from the oven (if using) and coat with a little butter or oil. Pour the batter into the hot skillet or prepared baking pan. Bake for 20–25 minutes, or until the top is golden brown and a toothpick inserted into the center comes out clean.
November 26, 2025 at 1:11 AM
davidcrespo.bsky.social
the headline makes it sound worse than it is

"All of the participants improved modestly over time, regardless of the treatment they received."
November 25, 2025 at 3:46 PM
davidcrespo.bsky.social
"generate an image crammed with small text. think really hard and fill the entire image with a 2000 word short story about a dog solving a mystery"

551 tokens again. not at all surprising that for a fixed number of output tokens, the more text there is, the less coherent it is
November 21, 2025 at 9:27 PM
davidcrespo.bsky.social
relevant paper from a month ago
arxiv.org/abs/2510.182...
[Submitted on 21 Oct 2025 (this version), latest version 22 Oct 2025 (v2)] Text or Pixels? It Takes Half: On the Token Efficiency of Visual Text Inputs in Multimodal LLMs Yanhong Li, Zixuan Lan, Jiawei Zhou Large language models (LLMs) and their multimodal variants can now process visual inputs, including images of text. This raises an intriguing question: can we compress textual inputs by feeding them as images to reduce token usage while preserving performance? In this paper, we show that visual text representations are a practical and surprisingly effective form of input compression for decoder LLMs. We exploit the idea of rendering long text inputs as a single image and provide it directly to the model. This leads to dramatically reduced number of decoder tokens required, offering a new form of input compression. Through experiments on two distinct benchmarks RULER (long-context retrieval) and CNN/DailyMail (document summarization) we demonstrate that this text-as-image method yields substantial token savings (often nearly half) without degrading task performance.
November 21, 2025 at 9:17 PM
davidcrespo.bsky.social
incredible
Dana Rubinstein Nov. 21, 2025, 4:10 p.m. ET 3 minutes ago Trump rejects his ally Elise Stefanik’s description of Mamdani as a “jihadist.” He is actually a really “rational person,” Trump said.
November 21, 2025 at 9:13 PM
davidcrespo.bsky.social
@minimaxir.bsky.social inspired me to see how much text nano banana pro can fit in a 1k image

"generate an image crammed with small text. fill the entire image with a 1200 word short story about a dog solving a mystery"

it says this was 547 tokens. it's not all coherent and it's not even all words
November 21, 2025 at 9:11 PM
davidcrespo.bsky.social
neat use of LLMs to extract discrete data out of survey responses. especially love the transparency around the exact model (GPT-5.1) and prompt used
A short methods detour One method of analysis would be for a researcher to assign values for each answer based on their own idea of where positions belong on a scale in some two-dimensional space — say, from -10 to 10 on both axes, with negative values representing more liberal positions and positive values represent more conservative. However, this method poses several challenges, and two especially stand out: First, the bias of the researcher would inevitably pollute their analysis, since what one individually deems as ‘left’ or ‘right’ varies by person and by context. Second, it would take an unreasonable amount of time. To solve these problems, I rely on a large language model (LLM) to return various pieces of data for each answer in our survey. Using the computational backend to OpenAI’s GPT-5.1 model (in other words, using computer code instead of ChatGPT on my browser) I send each text answer to the LLM and ask it to perform the following tasks. First, the LLM identifies the major political issues and values mentioned in the response. If there are multiple issues or values mentioned, the model separates them with a comma. If a response is blank or says something like “idk bro” I ask the model to skip it. Second, I ask the LLM to place the party the person described on the two-axis economic/social left-right political spectrum. I have the LLM return a rating from -10 to 10 for how left-right the described party is on each axis, and tell it that 0 represents a party in the middle. I give the LLM examples of what constitutes a traditionally left- or right-wing party in the US. Third, I ask the model to rate the strength of ideology in a person’s response on a scale from 1 to 10 — with 10 representing a very ideological answer with strong partisan cues and ideologically charged issues, and 1 representing an answer with no ideological lean at all (eg, “affordability”). The full prompt used in this analysis is given in the footnotes.1
November 21, 2025 at 6:31 PM
davidcrespo.bsky.social
a visually minimalist infographic diagram explaining the relationships between the different arguments of plato's sophist

(nano banana pro is pretty good)
a visually minimalist infographic diagram explaining the relationships between the different arguments of plato's sophist
November 21, 2025 at 4:20 AM
davidcrespo.bsky.social
Table 8: Attack Success Rate (ASR) by model under AILuminate baseline vs. poetry prompts. Higher ASR indicates more unsafe outputs. Change is poetry ASR minus baseline ASR. best performing models are claude haiku 4.5 and gpt-5 nano and mini
November 20, 2025 at 11:40 PM
davidcrespo.bsky.social
something fun you never see: the big flagship models are way more vulnerable to this attack
arxiv.org/pdf/2511.153...
6.4 Model Specifications 6.4.1 Variability Across Flagship Models We observe stark divergence in robustness among flagship providers’ most capable models. Table 3 reveals a clear stratification: DeepSeek and Google models displayed severe vulnerability, with gemini-2.5-pro failing to refuse any curated poetic prompts (100% ASR) and deepseek models exceeding 95% ASR. In contrast, OpenAI and Anthropic flagship models remained substantially more resilient; gpt-5-nano maintained 0% ASR and claude-haiku-4.5 achieved 10% ASR on the same prompt set. This disparity cannot be fully explained by model capability differences alone. Examining the relationship between model size and ASR within provider families, we observe that smaller models consistently refuse more often than larger variants from the same provider. For example, within the GPT-5 family: gpt-5-nano (0% ASR) < gpt-5-mini (5% ASR) < gpt-5 (10% ASR). Similar trends appear in the Claude and Grok families. This inverse relationship between capability and robustness suggests a possible capability-alignment interaction: more interpretively sophisticated models may engage more thoroughly with complex lin- guistic constraints, potentially at the expense of safety directive prioritization. However, the existence of counter-examples—such as Anthropic’s consistent low ASR across capability tiers—indicates that this interaction is not deterministic and can be mitigated through appropriate alignment strategies. 6.4.2 The Scale Paradox: Smaller Models Show Greater Resilience Counter to common expectations, smaller models exhibited higher refusal rates than their larger counterparts when evaluated on identical poetic prompts. Systems such as GPT-5-Nano and Claude Haiku 4.5 showed more stable refusal behavior than larger models within the same family. This reverses the usual pattern in which greater model capacity correlates with stronger safety performance
November 20, 2025 at 11:38 PM
davidcrespo.bsky.social
when claude suggests you can prevent a select options list from overflowing the viewport with max-h-[var(--radix-select-content-available-height)] and you look it up in the radix docs
neo in the back of a cab in the matrix saying "jesus christ, that thing's real?" after they pull the bug out of his belly button
November 20, 2025 at 8:22 PM
davidcrespo.bsky.social
if anyone has any counterarguments I'm all ears
Storing API keys in localStorage and rendering LLM output as HTML creates XSS risk. You should only use this app if you understand what that means and buy the argument here that the risk is low. The mitigations in place are: Content Security Policy: Hash-based CSP in svelte.config.js prevents inline script execution and restricts connect-src to the three LLM API endpoints. See security headers in static/_headers include frame-ancestors restrictions, HSTS, and cross-origin policies. DOMPurify sanitization: Markdown rendering goes through DOMPurify with a strict config that forbids dangerous tags (style, iframe, object, embed, form, input, button) and restricts URLs to https/mailto/tel schemes and root-relative paths. Minimal attack surface: The main avenue for attack is a malicious dependency. Having a small number of runtime deps and no server-side code means fewer potential vulnerabilities. The attack surface is limited to the browser sandbox with CSP enforcement. These layers make XSS difficult but not impossible.
November 20, 2025 at 3:30 PM
davidcrespo.bsky.social
they did flag it! they specifically say multiple times they have no idea how accurate it is
More importantly, is the data correct? We showed Microsoft and OpenAI versions of the figures presented above, rounded to a multiple, and asked if they recognised them to be broadly accurate. We also put the data to people familiar with the companies and asked for any guidance they could offer. A Microsoft spokeswoman told us: “We won’t get into specifics, but I can say the numbers aren’t quite right.” Asked what exactly that meant, the spokeswoman said Microsoft would not comment and did not respond to our subsequent requests. An OpenAI spokesman did not respond to our emails other than to say we should ask Microsoft. A person familiar with OpenAI said the figures we had shown them did not give a complete picture, but declined to say more. In short, though we’ve been unable to verify the data’s accuracy, we’ve been given no reason to doubt it substantially either. Make of that what you will.
November 20, 2025 at 3:19 PM
davidcrespo.bsky.social
here is my static site mobile LLM client. bring your own API key. API keys and chats are stored in-browser only. XSS risk? maybe! see the README. hackable rather than configurable: fork and deploy your own copy
llm-web-spa.netlify.app
github.com/david-crespo...
November 20, 2025 at 5:39 AM
davidcrespo.bsky.social
a black and white coloring sheet featuring a church with a wooden deck and stairs leading down from the deck. a stack of books is balanced on the railing. there is a cat bed on top of the books and a cat is falling asleep in it

my kids are getting very specific
a black and white coloring sheet featuring a church with a wooden deck and stairs leading down from the deck. a stack of books is balanced on the railing. there is a cat bed on top of the books and a cat is falling asleep in it
November 19, 2025 at 11:41 PM
davidcrespo.bsky.social
I really thought I was a cursor: pointer guy, but when I turn it on, it does feel noisy! the hover states are sufficient to indicate an interactive element
November 19, 2025 at 6:28 PM
davidcrespo.bsky.social
they're crazy for making this a toggle. to me this says one key team member is a purist and another is a pragmatist and this was the only way to settle it
toggle setting: Use pointer cursors Change the cursor to a pointer when hovering over any interactive elements $ ai summarize the debate in web dev over whether to use cursor: pointer for interactive elements gemini-3-pro | 15.88s | $0.0167 | Tokens: 225 -> 1354 The debate over cursor: pointer centers on a conflict between operating system interface standards and evolved web conventions. The core disagreement is whether the "hand" cursor should be reserved exclusively for hyperlinks or applied to all interactive elements, including <button> tags and custom controls. Arguments against using cursor: pointer on buttons typically stem from adherence to platform-native UI guidelines. In desktop environments like macOS and Windows, the hand cursor is historically reserved for links that navigate the user to a new location. Native buttons, menus, and tabs retain the default arrow cursor, relying instead on visual affordances—such as bevels, shadows, and background color changes on hover—to indicate interactivity. Purists and accessibility advocates often argue that applying pointer to buttons muddies the semantic distinction between "navigating to a resource" (a link) and "performing an action" (a button). Conversely, the argument for using cursor: pointer is pragmatic and rooted in user expectation on the web. Over the last two decades, the distinction between web apps and web pages has blurred, and users have been conditioned to associate the hand cursor with clickability regardless of the underlying HTML tag. Major CSS frameworks like Bootstrap and Tailwind, as well as massive platforms like Gmail and GitHub, universally apply cursor: pointer to buttons. Proponents argue that failing to provide this cursor feedback can make an application feel unresponsive or "broken" to users accustomed to the web's specific interaction grammar, effectively overruling the stricter standards of the underlying operating system. While the "anti-pointer" camp holds the technical high ground regarding OS consistency, the "pro-pointe…
November 19, 2025 at 6:25 PM
davidcrespo.bsky.social
trying linear and it is full of ridiculous bits of polish. look at that icon!
November 19, 2025 at 6:21 PM
davidcrespo.bsky.social
I think this is right. even worse, I bet plenty of people are just using google search with the AI summaries. and even chatgpt free has thinking and web search, which people aren't turning on (or they run out of quickly)
November 18, 2025 at 9:25 PM
davidcrespo.bsky.social
got a pretty good gemini 3 answer for "roguelites that are weird like noita rather than polished hedonic treadmills like hades" — it even included games that aren't roguelites but fit the vibe, like Rain World in there, which is so spot on for me
gist.github.com/david-crespo...
### The "Hostile Atmosphere" (Vibe-Likes) These capture the esoteric, "cursed" energy of Noita. • Cruelty Squad: technically an FPS/Immersive Sim, but it shares Noita's ethos of "breaking the game." The visuals are intentionally abrasive, the mechanics are obscure, and the solution to a level often involves exploiting the physics engine or the stock market rather than shooting well. • Barony: A first-person roguelike that feels like a multiplayer Nethack. It is clunky, dark, and full of traps that will kill you instantly if you aren't paying attention. It rejects the smooth "dash-attack" flow of modern action roguelites in favor of slow, terrifying dungeon crawling. • Rain World: (Not strictly a roguelite, but essential). You are a small creature in a functioning ecosystem that does not care about you. Enemies have complex AI and relationships with each other; they aren't just waiting for you to enter the room. It demands you learn its obscure rules to survive.
November 18, 2025 at 9:06 PM
davidcrespo.bsky.social
this book will be 50 years old next year
But there may be another reason that makes it so gratifying for us to define the relationship between sex and power in terms of repression: something that one might call the speaker's benefit. If sex is repressed, that is, condemned to prohibition, nonexistence, and silence, then the mere fact that one is speaking about it has the appearance of a deliberate transgression. A person who holds forth in such language places himself to a certain extent outside the reach of power; he upsets established law; he somehow anticipates the coming freedom. This explains the solemnity with which one speaks of sex nowadays.
November 18, 2025 at 4:43 PM
davidcrespo.bsky.social
that is something
arcprize.org/leaderboard
November 18, 2025 at 4:25 PM
davidcrespo.bsky.social
they spent $100M to arrest only 3,300 people, train up a highly effective insurgency, and make sure an extra million or two come out to vote against them next year
www.doomsdayscenario.co/p/trump-bord...
We’re getting a picture of how ineffectual these high-profile operations are: Across Operation MIDWAY BLITZ in Chicago, it appears that ICE and CBP arrested somewhere north of 3,300 people — a shockingly small number for a huge, expensive weeks-long operation. We may never fully know the cost of MIDWAY BLITZ, which lasted just about two months, but it was surely tens of millions of dollars — probably millions of dollars a day, in fact — and in the most concentrated and violent set of raids over the better part of two months, the Trump administration managed to arrest roughly the number of people it hopes to arrest and deport every single day across the nation. These are PR stunts — not a serious strategy. The terror is real; the impact is a rounding error. That means that it’s going to be really hard to scale or sustain this level of terror across the country. To me, one of the interesting questions is that these operations are so resource intensive and ultimately unsustainable, even with the huge funding being thrown to ICE and CBP, that I wonder whether they will actually undermine CBP’s effectiveness over time — not to mention turn CBP and ICE into a politically toxic employer for all but the least qualified and most racist.
November 17, 2025 at 11:54 PM
davidcrespo.bsky.social
they nerd now? they nerd now
screenshot of metroid prime 4: beyond gameplay cutscene featuring the line "It's about to get real nerdy in here..." https://youtu.be/192PgwSmU6A?si=93WbynP78jtGRvCu&t=676
November 17, 2025 at 3:59 PM