Source: dbushell.com
Author: David Bushell
Date: 2026-04-29
"GitHub used to be cool and now it's a lame slop graveyard."
Under Microsoft, GitHub has deteriorated into an unreliable, "enshittified" platform flooded with AI slop and bots. The author argues it is now an expensive liability rather than a useful tool, and urges users to begin migrating immediately.
Source: github.com/oobabooga/textgen
Author: oobabooga
Date: Active development (5,600+ commits, 112 releases)
TextGen is an open-source desktop application designed to run large language models (LLMs) locally on consumer hardware. It provides both a privacy-first web interface and an OpenAI/Anthropic-compatible API, enabling users to self-host chat, vision, tool-calling, and web-search capabilities without relying on third-party cloud services or sending data externally.
instruct, chat-instruct, and chat modes.diffusers, and extensions for TTS, voice input, and translation.Download a pre-built portable release from GitHub Releases. Includes all dependencies (CUDA, Vulkan, ROCm, CPU-only). Compatible only with single-file GGUF models.
Run the OS-specific start script (start_windows.bat, start_linux.sh, start_macos.sh). Creates a local environment and launches the UI at http://127.0.0.1:7860. Supports environment variables for silent/automated installs.
For users needing multi-file model support (Transformers, EXL3), training, or extensions:
- Requires Conda (Miniforge recommended) and Python 3.13
- Install PyTorch 2.9.1 with hardware-specific wheels (NVIDIA CUDA 12.8, AMD ROCm, Apple MPS, CPU-only)
- Clone repo and install from requirements/full/
- Requires ~10GB disk space
| Format | File Structure | Portable Build | Full Install |
|---|---|---|---|
| GGUF | Single .gguf file in user_data/models |
✅ Yes | ✅ Yes |
| Transformers | Multi-file folder in user_data/models/<name>/ |
❌ No | ✅ Yes |
| EXL3 | Multi-file folder | ❌ No | ✅ Yes |
Users are directed to resources like LocalBench for recommended GGUF quantizations and a VRAM calculator for memory planning.
TextGen is explicitly built for users who want full control over their AI workloads. By running locally, users avoid data leakage, API rate limits, and subscription costs. The project emphasizes ease of use for beginners while offering deep configurability for advanced users through command-line flags, persistent settings, and an extension ecosystem.
Source: github.com/nathanlgabriel/paper_code_mapping_assessment
Author: Nathan L. Gabriel
Date: 2025-07-01
Local LLMs have advanced dramatically enough to perform a task previously impossible for small local models: mapping computational simulation code to its corresponding academic research paper with high accuracy. However, the more surprising finding is that a decent local model combined with an intelligent human can outperform even overpowered frontier models.
Evaluate local LLMs on their ability to map computational simulation code to its corresponding research paper, tracking iterative refinement from initial outputs to corrected mappings.
"Ultimately, my assessment is that the hype is real. Qwen 3.6, Gemma 4, and Nemotron Nano were all able to do reasonably well at a task that was impossible for small local models a few months ago."
The author discovered a substantive oversight in Claude Sonnet's mapping of replicator dynamics related to continuous vs. discrete population representation and rounding bias prevention.
"Conclusion: A decent local model + an intelligent human can still be smarter than an overpowered frontier model."
| Model | Performance |
|---|---|
| Qwen 3.6 35B A3B | Standout performer. Captured 75-80% of the definitive mapping. Fast inference. |
| Qwen 3.6 27B | Reasonable baseline; solid structural understanding but required more extensive correction. |
| Gemma 4 26B | Generated initial mappings requiring substantial correction; good grasp of simulation pipeline. |
| Nemotron Nano | Delivered reasonable baseline outputs. |
| Gemma 4 31B | Failed: Exceeded context/VRAM limits; resulted in inference crashes. |
The hype around local LLMs is justified — they can now handle complex code-to-paper mapping tasks that were impossible months ago. But the more profound insight is about human-AI collaboration: careful prompting and human oversight with a capable local model can catch errors that even the most powerful frontier models miss.
Source: LightOn Blog
Date: 2026-05-12
Agent-ModernColBERT is a 149M-parameter late-interaction retriever that adds ~10% accuracy over Reason-ModernColBERT on BrowseComp-Plus by treating agent reasoning traces as first-class retrieval signals. Paired with GPT-OSS-120B, it matches the original GPT-5 + Qwen3-Embed-8B stack while being 54× smaller than dense alternatives.
Source: arXiv:2605.10938
Authors: Keya Hu*, Linlu Qiu*, Yiyang Lu, Hanhong Zhao, Tianhong Li, Yoon Kim, Jacob Andreas, Kaiming He (MIT) · equal contribution
Code:* github.com/lillian039/ELF
Date: 2026-05-11
Continuous diffusion models can beat discrete ones for language — if you stay in embedding space until the very last moment. ELF achieves Gen. PPL ~24 with just 32 sampling steps, using 10× fewer training tokens than leading discrete DLMs, with no distillation required.
Source: paulgraham.com
Author: Paul Graham
Date: July 2023
Graham collected techniques for doing great work across many fields and found their intersection. The result is a definite shape — not just a point labeled "work hard." The essay is a recipe for the very ambitious, organized around four steps: choose a field, reach the frontier, notice gaps, explore them.
The work must have three qualities: natural aptitude, deep interest, and scope for great work. The third is rarely a problem for ambitious people.
How to find it: By working. If unsure, guess and start. Develop a habit of working on your own projects — great work almost always happens on projects you drive yourself.
"What are you excessively curious about — curious to a degree that would bore most other people? That's what you're looking for."
The Four Steps:
Steps two and four require hard work. Interest drives harder work than diligence ever could. The three most powerful motives: curiosity, delight, and the desire to do something impressive. When they converge, the combination is the most powerful of all.
You can't know what a field is like without doing it — so the four steps overlap. You may work at something for years before knowing if it's right. Ambition comes in two forms: one that precedes interest (makes deciding harder) and one that grows out of it.
Educational systems pretend choosing is easy and expect early commitment. In reality: "When it comes to figuring out what to work on, you're on your own."
Be a big target for luck: try many things, meet many people, read many books, ask many questions. When in doubt, optimize for interestingness. A field should become increasingly interesting as you learn more. If it doesn't, it's probably not for you.
Signs of fit: you enjoy even the parts others find tedious or frightening. Strange tastes are often strong ones — and strong taste means you'll be productive.
Make what you yourself want. Write the story you want to read; build the tool you want to use. Many people get this wrong, trying to please an imaginary sophisticated audience.
Forces that lead you astray: pretentiousness, fashion, fear, money, politics, other people's wishes, eminent frauds. If you stick to genuine interest, you're proof against all of them.
Following genuine interests requires boldness — risking rejection and failure. But you don't usually need much planning. The recipe is: work hard on excitingly ambitious projects, and something good will come of it.
Planning only works for achievements you can describe in advance (gold medals, wealth). You can't plan discovery of natural selection. Instead, "stay upwind": at each stage do what seems most interesting and gives the best options for the future.
Work hard, but not too hard. For the hardest types, you may only manage 4–5 focused hours a day. Arrange large contiguous blocks — you'll shy away from hard tasks if interruption is likely.
Activation energy: Starting is harder than continuing. Trick yourself: "I'll just read over what I've got so far." Similar lies work for starting projects: "How hard could it be?" The young have an advantage here — their optimism (sometimes born of ignorance) helps overcome activation energy.
Finish what you start. Much of the best work happens in what was meant to be the final stage. Exaggerating the importance of your work (in your own mind) is another permissible lie — if it helps you discover something new, it may turn out not to have been a lie after all.
Two forms of procrastination: - Per-day procrastination — obvious, sets off alarms - Per-project procrastination — far more dangerous. It camouflages itself as work: you're industriously busy on something else. Ask regularly: "Am I working on what I most want to work on?"
Consistency is king. Writing a page a day = a book a year. People who do great things don't get a lot done every day — they get something done, rather than nothing. We underestimate cumulative effects. Exponential growth (learning, audience) feels flat early on — push through the initial unrewarding phase.
Undirected thinking: Letting your mind wander (while walking, showering) solves hard problems, but only when interleaved with focused work that feeds it questions. Avoid distractions that push your work out of the top spot in your mind. (Exception: Don't avoid love.)
Aim high. Cultivate taste — know the best in your field and what makes it so. Trying to be the best simplifies things and is often easier than merely trying to be good. Use the 100-year test: will people care in a century?
Style emerges from doing your best work; don't try to be distinctive. Trying to be is affectation — pretending someone other than you is doing the work. The fakeness shows.
"Be the one who puts things out there rather than the one who sits back and offers sophisticated-sounding criticisms."
Earnestness is the positive expression: intellectual honesty, willingness to admit error, informality. The core is intellectual honesty — to see new ideas you need an exceptionally sharp eye for the truth. Maintain a slight positive pressure toward admitting you're wrong. Once you admit a mistake, you're free. Till then you carry it.
Nerds have an advantage: they expend little effort on seeming anything. Any energy that goes into how you seem comes out of being good.
Great work is consistent with itself. When facing a decision, ask: "If I'd already made the change, would I want to revert to what I have now?"
Have the confidence to cut. Don't keep something that doesn't fit just because you're proud of it or it cost effort.
Elegance is a useful standard beyond math. Laborious solutions often have more short-term prestige — they cost effort and are hard to understand. But the best work sometimes seems like it took little effort, because it was in a sense already there, just waiting to be seen.
When you're doing work that could be seen as either creation or discovery, err on the side of discovery. Think of yourself as a conduit through which ideas take their natural shape.
Make tools gratuitously unrestrictive — a powerful tool will be used in ways you didn't expect. Express ideas in the most general form; they'll be truer than you intended.
Originality is not a process — it's a habit of mind. Original thinkers throw off new ideas about whatever they focus on, like an angle grinder throwing off sparks. They can't help it.
Ideas come from working on something slightly too difficult — not from trying to have original ideas. Writing is a powerful generator: when you try to put ideas into words, a missing idea creates a vacuum that draws it out of you. There's a kind of thinking that can only be done by writing.
Changing context helps: visiting new places, going for a walk. Also travel in topic space — explore different topics. Distribute attention according to a power law: be professionally curious about a few things and idly curious about many more.
New ideas are simultaneously novel and obvious. The contradiction resolves because seeing them requires changing your model of the world. Models both help and constrain us. To find new ideas, be stricter than others — notice where models bash against reality. This is what Einstein did with Maxwell's equations.
Rule-breaking is required because new models break at least implicit rules of old ones. Two ways to be comfortable breaking rules: aggressive independence (enjoying it) and passive independence (not caring or not knowing the rules exist). Novices and outsiders often make discoveries because their ignorance acts as passive independent-mindedness.
"A good new idea has to seem bad to most people, or someone would have already explored it."
Look for ideas that seem crazy but the right kind of crazy — they tend to be exciting and rich in implications.
Choosing problems > solving problems. People show more originality in solving than in deciding which problems to solve. Unfashionable problems are undervalued. One of the most interesting kinds: problems people think have been fully explored, but haven't. Ask yourself: if you took a break from "serious" work to work on something purely because it would be interesting, what would you do? The answer is probably more important than it seems.
Questions matter more than we think. People imagine big ideas are answers, but often the real insight was in the question. A really good question is a partial discovery. Be rich in unanswered questions — they grow in the answering.
Being prolific is underrated. Try lots of things — you can't have many good ideas without also having many bad ones. Err on the side of starting. Great things are made in successive versions: start small and evolve.
Begin with the simplest thing that could possibly work. Surprisingly often, it does. Don't cram too much into one version. An early version being dismissed as a "toy" is a good sign — it has everything a new idea needs except scale, and that tends to follow.
Planning is a necessary evil — a response to inflexible media or coordinating many people. If you keep projects small and use flexible media, designs can evolve instead.
Take as much risk as you can afford. In an efficient market, risk is proportionate to reward. If you're not failing occasionally, you're probably being too conservative.
Advantages of youth: energy, time, optimism, freedom. Advantages of age: knowledge, efficiency, money, power. The young often don't realize how rich they are in time. Spend it lavishly but not wastefully — learn something out of curiosity, build something because it's cool, become freakishly good at something.
Fresh eyes: When learning something for the first time, pay attention to things that seem wrong or missing. There's a 99% chance the problem is with you, but if the misgiving survives as your knowledge grows, it may represent an undiscovered idea.
Unlearn school-induced nonsense: Schools induce passivity (authority tells you what to learn and tests you), give a misleading impression of work (problems are pre-defined and always soluble), and train you to win by hacking the test. You can't trick God. Stop looking for shortcuts — focus on problems others have overlooked.
Don't take rejection by committees to heart. The qualities that impress admissions officers differ from those required to do great work.
There's a good way to copy and a bad way. Copy openly, not furtively or unconsciously. Originality is the presence of new ideas, not the absence of old ones. Early work is almost inevitably based on others' — you have no previous work of your own to react to yet.
Don't copy the flaws. The features easiest to imitate are most likely the flaws. Some talented people are jerks — being a jerk is not part of being talented, it's merely how they get away with it.
Cross-pollinate by copying ideas from one field into another. Negative examples can be as inspiring as positive ones — sometimes you learn more from things done badly.
Seek the best colleagues. Quality over quantity — one or two great ones beat a building full of pretty good ones. Work with people you want to become like, because you will. Sufficiently good colleagues offer surprising insights — they can see and do things you can't.
Husband your morale like a living organism. Morale compounds: high morale → good work → higher morale → even better work. The cycle also runs in reverse. When stuck, switch to easier work just to get something done.
Treat setbacks as part of the process. Solving hard problems always involves backtracking. Never let setbacks panic you into backtracking more than necessary. Learn to distinguish good pain (effort) from bad pain (damage).
An audience matters for morale — but it doesn't need to be big. A handful of people who genuinely love what you're doing is enough. Avoid intermediaries between you and your audience whenever possible.
People affect your energy. Seek those who increase it, avoid those who drain it. Don't marry someone who doesn't understand that you need to work.
Morale is physical. Exercise, eat and sleep well, avoid dangerous drugs. Running and walking are particularly good for thinking.
People who do great work aren't necessarily happier than everyone else — but they're happier than they'd be if they didn't. For the smart and ambitious, not being productive is dangerous — it breeds bitterness.
Fame adds noise. The opinion of people you respect is signal. The prestige of a type of work is at best a trailing indicator. The right question: how well could it be done? not how much prestige does it have?
Competition can motivate, but don't let it choose your problem for you. Don't let competitors make you do anything more specific than work harder.
The factors in doing great work are literally mathematical: ability, interest, effort, and luck. Luck you can't control, and effort is assumed if you genuinely want to do great work. So it boils down to ability and interest combining to yield new ideas.
Curiosity is the word that appears most often. It's the key to all four steps: it chooses the field, gets you to the frontier, makes you notice the gaps, and drives you to explore them. The whole process is a kind of dance with curiosity.
The essay's length itself acts as a filter — if you made it this far, you're already further along than you realize. Many more people could try to do great work than do. What holds them back is a combination of modesty and fear.
"The discoveries are out there, waiting to be made. Why not by you?"
Source: sockpuppet.org
Author: tqbf (Thomas Ptacek)
Date: 2026-05-12
AI agents have made bespoke native UI development so accessible that software culture is becoming "Emacsified" — hyper-personal, idea-driven, and endlessly configurable, with native interfaces replacing both clanky TUIs and bloated Electron apps.
The author argues that AI agents have triggered a "cursed renaissance of TUI tooling," making the Markdown reading experience intolerable.
glow, Markless): Well-built but "hamstrung by the terminal itself," which is almost always monospaced and fatiguing..md file disrupts carefully arranged editing environments."I'm certain that at least 14% of the agita about AI code is driven by exhaustion over incessantly scrolling terminal Markdown."
Instead of settling, the author had Claude generate a native macOS Markdown viewer: MDV.app.
"MDV isn't the best macOS application ever built... But it has improved my quality of life immensely."
The author urges readers not to install it, but to treat it like a shiny .emacs: "Steal the idea and make a better one."
Signal is cited as an Electron app ("a whole-ass copy of Chromium") causing constant screen flickering. Historically, native UI was rare because capable macOS developers are "rare birds."
"Claude isn't just a replacement-level SwiftUI developer. Claude is actually good."
A new software culture where AI agents bypass the traditional pain of native UI development, allowing Emacs-like bespoke tooling to escape into the mainstream.
.emacs.""For the kinds of software I'm talking about, you want the prompts more than you want the source code."
"'Building' implies more effort than you're expending. What you're doing feels a lot more like configuring, on a platform that has suddenly become vastly more configurable. A platform that feels a lot more like Emacs."
"AI agents have fracked Emacs culture, and it's leaking out into the wider world."
"Building native UI is now fun; a lot more fun than building web interfaces ever was. Give it a shot; make something stupidly specific to your own problems, enjoy it for a little while, and then throw it away and make something else."