TrendPulse

Hey, I'm a former Philosophy lecturer and I have essentially zero experience with llms or how they work. Having said that, much of my adult life has been devoted (wisely or not) to offering up definitions of a variety of phenomena. This sub is a brilliant space for challenging my concepts and prejudices. For example, while reading the post about whether or not a submarine can swim, I found myself flip-flopping about my understanding of what it means "to think". It reminds me of...

I am subbed to various AI-related subs, but I really want to find people doing deep dives on this specifically. Any pointers will be much appreciated. Cheers.

A year ago a decent executor model cost real money per million tokens. Now there are three or four sitting in the near-free tier. On paper that's a collapse. In practice my spend is flat, and the reason is boring. A cheaper model gets re-run. It stops early, or it silently substitutes a literal path where a glob was supposed to go, or it burns most of its output budget thinking before it says anything, and the orchestrator sends it around again. Three cheap attempts at a step is not cheaper th...

Hugging Face Artificial Analysis Should be a sweet spot for general work. Seems like coding is the only part that is inferior to Qwen.

https://preview.redd.it/83kvawwdu7hh1.png?width=2520&format=png&auto=webp&s=59709d86f3b6d7f599acbd88fa3693a5e49653f3 Cheap Opus 4.7 replacement. Good job to qwen team.

Yesterday I saw a random video about a guy that used chatgpt to basically create farming software for himself rather then paying 10s of thousands of dollars for it. Though we are not in post scarcity for physical products it seems like we are getting there for useful software.

Today I learned that ChatGPT can easily generate a watch face with a specified time. A year ago, this used to be a classic AI fail. You ask for any time, and it draws 10:10, no matter what. Because most of the internet stock photos showed 10:10, and the frontier models majorly learn through imitation learning. It seems OpenAI/Anthropic has specifically trained the model for this use case. What other "10:10 problems" are still out there? Stuff that fails not because the model can...

It's not impossible for Google to compete with openai or Anthropic. They have the resources and money. But there's a trade because of how fucking expensive and resource intensive ai development is. Google, as a company (which the point of companies is to make capital), does not depend on having the top frontier or SOTA models as much as you guys hope. Actually, the top 10 and over don't

Hi there, I'm a software engineer, using LLMs heavily both in my day job and for pet projects, and I started noticing some signs, that are a bit worrying me. Maybe even more than a bit. There are already a lot of claims of how LLMs are addictive, creating non-stop flow of dophamine, as we see how they help us research or build something - and yes, I've noticed it myself. There are also claims that working with LLMs is akin to playing with slot machine - and for pure vibe-coding it is definitel...

https://arxiv.org/pdf/2607.28607 Aligning large language models to prevent them attributing consciousness to themselves inadvertently alters their representations of mindedness in other entities alongside human beliefs and values. We demonstrate that safety fine-tuning suppresses models' tendencies to attribute minds not only to themselves, but also to non-human animals and natural objects, while also driving a reduction in spiritual belief. Both ablating the learned safety-refusal directio...

https://x.com/i/status/2084319326895804869 We train humanoid robots to vault obstacles, climb platforms, and cross terrain they’ve never seen—all from a single perceptive, whole-body policy. It decides whether to walk, climb, or vault, and switches between skills on its own. No skill labels. No state machines. No motion generators at inference.

The voluntary framework outlined in the June 2nd executive order is complete. Discussions with industry about next steps are underway. Reps from the AI companies will be meeting at WH tomorrow to discuss and get a look. It's unclear whether the framework is now in effect. Companies were lobbying for specific language on issues including open-source as late as last week. Source: The Information

https://x.com/intology/status/2084319121332965804/photo/1 These results are from Intology: https://x.com/intology/status/2084319121332965804 Their Locus system post-trained qwen3 base models beyond the qwen3 instruct checkpoint following the PostTrainBench setting: https://posttrainbench.com/

Seeing OpenAI's recent progress surprised me, but what surprised me even more is how little Google DeepMind appears in comparison. DeepMind has arguably done more foundational work in AI for mathematics than anyone else over the past few years, they got AlphaGeometry, AlphaProof, AlphaEvolve and a long history of research on reasoning and search. If you'd asked me a 1-2 years ago which lab would dominate difficult mathematical benchmarks, I probably would have said DeepMind. Is this simply b...

my take: it's true that the ratio of code written to code reviewed has shifted as engineers started to spend more time reviewing code than writing it. no surprise there. but we stopped writing assembly because compilers provide a "deterministic", reproducible translation. we can inspect the generated assembly when something goes wrong. ai generated binaries have neither property, i'd assume - if this is what he's talking about. so the mapping is stochastic, and the output isn't meani...

Companies like JPMorgan, Morgan Stanley, Walmart, Uber, and Salesforce have the capital, infrastructure, and technical talent to run open weight models themselves. As these models improve, large enterprises no longer need to pay a premium for access to closed models. They can own the weights, customize the models with proprietary data, control where their data goes, and avoid dependence on one provider. So if they loss say 50 percent of their enterprise clients in the next 24 months where do...

I have intentionally compared the performance of this model to other models in its weight class (200-300B parameters). Check out the results for yourself here: Inkling Small - Intelligence, Performance & Price Analysis

128gb of unified memory is somewhere north of four grand right now. A subscription that covers what I actually do is twenty a month. That's a payback period measured in decades and the machine will be a paperweight in five years. I keep doing this arithmetic, keep getting the same answer, and keep wanting the answer to be different. So here's why I still think about it and someone can tell me which part is cope. Privacy is real but it isn't four thousand dollars real for what I do. Offline i...

I've run this harness in GPT 5.6-Sol Pro for 679 minutes total (11 hours, 19 minutes) A few failures, and two discoveries. One was a very niche problem that had like a few papers on it (it improved the bound) Then I re-prompted it, to only consider problems that at least have a dedicated Wikipedia page. It autonomously scans, use the theorem prover it wrote in C++, reads the relevant papers, and boom. New record. Full convo: https://chatgpt.com/share/6a6c9582-2a58-83ee-8123-c9a90a7657...