if you've been using latest frontier LLMs, it's almost certain that you would have noticed by now the newer models have become worse to talk to they're more robotic, they speak jargons, they spits out verbose text, and do stuff you didn't ask for how did that happen? well, i'm not the person who trained those models so i can't speak for certain, but i've known enough evidence that gives me a well-educated guess, and i thought it's interesting to share as a crash course of modern LLM training pipelines so here we go let's wind back to 2020. GPT-2 and GPT-3 already came out and were widely available, but they could only predict one token at a time - that's what LLMs are at their core token prediction was offered via API, but there was nothing you could "talk to". so while it generated a lot of excitement in the academic field due to the emergent intelligence, it didn't have any wide adoption in 2022, ChatGPT changed all that. the research work that led to ChatGPT was a model initially named "InstructGPT". it took GPT-3 as the intelligent base, and used reinforcement learning with human feedback (RLHF) to teach the models how to "chat" the core idea of RLHF is that you ask the model to generate a few responses, and then let real humans pick which one they like. do this over and over again, and you get a model that knows how to talk worth noting even as early as InstructGPT, research found that making the model more pleasant to talk to will reduce their pure academic capabilities. this was called "alignment tax", which is an interesting thing we'll come back to in a bit there were various techniques done to minimize the reliance on humans, but ultimately the reward is modeled after human preference, making these AI assistants easy to talk to so remember this: RLHF = training the model to be likable by humans in 2024, there was an inflection point introduced by claude sonnet 3.5 which was the first model that can kind of autonomously finish coding tasks. it led to the first wave of viable "coding agents" the way sonnet 3.5 achieved this was by training the model with a harness (now it's called an agent) that has bash and file editing tools, throw the agent into a virtual machine, give it a task, and let it try to complete it. these tasks all have a machine-verifiable outcome predefined, mostly via test cases, that can validate whether the model really finished the task or not then you let the model do billions and billions of attempts in such virtual environments, and some of them would succeed by chance. you keep the successful agent sessions, and use reinforcement learning to teach the model to do that more, and boom - you get a coding agent that is called reinforcement learning with verifiable rewards (RLVR). if you look closely, you'll see that in this RLVR process, the final text response from the model doesn't matter AT ALL, as long as the code written by the agent could pass the test. it could talk like a jerk and it would still be rewarded so remember this: RLVR = training the model to be accepted by machines late 2024 and early 2025, we saw o1 and deepseek R1 came out as the first wave of "reasoning models". this article is getting long so i'm not diving into reasoning models now, but just know that reasoning models also relied heavily on RLVR to scale the training process - let the model think before taking action, and if the thinking led to a machine verifiable outcome, reward the thinking trace and teach the model to think like that more often the biggest difference between RLVR and RLHF is that RLVR is more scalable. human feedback is expensive to get, especially in domains where only an expert can have a valid opinion on which result is good with RLHF, if we let the model generate 100 responses, then a human has to review all 100 responses to pick which is good with RLVR, the human (or sometimes an AI) would define a task and verifier only once, and the model can generate a million responses - the machine verifier will pick which responses are good in an automated way so as a result, RLVR is becoming more and more dominant in newer models' training pipeline if you put all these things together: - RLHF = training the model to be likable by humans - RLVR = training the model to be accepted by machines - RLVR is more scalable - "alignment tax" says "likable by humans" makes the model do worse on verifiable tasks now you see why the newer models are becoming less and less likable? this is not just a "frontier labs screwed up their model training" problem - this is a war between machines and humanity, and humanity is losing we chased after benchmarks, when none of the benchmarks measure whether humans actually enjoy working with the model we use machines to decide which AI response is better because that's easier and cheaper, when we have no way of making sure those machines actually represent what we humans want we let AI go dark in a virtual environment on its own and complete predefined tasks at all costs, when in reality we often cannot define a verifiable outcome upfront, and need AI to work with us along the way i don't have a good solution to this, but i want to call for awareness that we're starting to witness a failure in aligning super intelligence right in front of our eyes this war between machines vs humanity is one we really can't afford to lose

many people ask when to use a big model (like sol/fable) at low reasoning effort, vs a small model (like luna/sonnet) at high reasoning effort i deliberately forced myself to use all the permutations a lot over the last couple of weeks to build intuition, and i realized the difference is "wisdom" vs "diligence" bigger models are "wiser" they have seen a lot. they remember a lot. they have a lot of expertise across different domains. they have better intuition, can connects the dots, and come up with creative, inspired ideas reasoning effort makes a model more "diligent" it'll assess each option, think through consequences, and figure out edge cases etc more thoroughly if there are 100 paths ahead, diligence makes the model assess every single one without a miss, but it will not make the model realize maybe the best one is to take none of the 100 paths and instead dig a tunnel "wisdom" and "diligence" are orthogonal. and now i get why the models are launched the way they were, and not just a single fable level model with 12 different reasoning levels we're all misled by the way we've been plotting the models with benchmark scores, which are fundamentally flawed because they use a single dimension to measure the model's capability, making us think of model size and reasoning effort as being fungible with each other, while in fact "wisdom" and "diligence" needs to be measured separately i hope the evals eventually catch up and address this. until then, here's my recommendation for how to choose - - if the problem you are trying to solve is something you think requires a genius, use a bigger model - if the problem you are trying to solve is something you think requires a pen and lots of paper, use higher reasoning effort - if it requires a genius sitting down with a pen and lots of paper, tune up both

opus 5 is a VERY interesting release for a few reasons 1. it showed that the general benchmarks we use today are almost completely useless now opus 5 is nowhere near fable in practical use, not even close. anyone who’s used it meaningfully can tell this very quickly after a few tasks. yet opus beats fable on many benchmarks i now trust domain specific benchmarks built with private datasets a lot more than the popular ones. perhaps the future is everyone running their own evals because the public ones are really not telling us much 2. it seems with the 5 series, anthropic is trying a new way of training models previously, the same generation of sonnet and opus were often released at the same time or sonnet comes out before opus, which indicates sonnet and opus were trained by separate pipelines in parallel with the 5 series, it was very clear that they trained mythos first, and then distilled it into sonnet and opus. it seems this approach has a big influence on the models seeing sonnet 5 being a flop and opus 5 getting pretty mixed reviews already, i’m not sure this is working out 3. “how pleasant is it to work with the model” used to be a strength in claude, but now it’s not. honestly, grok is my favorite right now on the “pleasant” dimension. kimi is not bad either it feels like both anthropic and openai are giving RLHF less care, in favor of scalable RL that’s machine verifiable this almost looks like AI is directing humans to build a world that’s more friendly for machines rather than humans, and most humans don’t even realize they are being manipulated to help with that almost every new generation of frontier models now talk more jargons, need more steering to do what you want, and are just less fun to work with if this continues, AI will start to speak their own language that looks like English but average humans can’t understand. they will choose to do things that their human user never asked for. are we already failing at alignment?

maybe this is controversial, but i believe what Cursor shipped here is a wrong solution to routing intelligence more generally, any attempt to do model routing at request level, while may yield some small gains, is fundamentally flawed and doomed to fail here's why - the complexity of a task only reveals itself when you start working on it. this is the same reason why we humans are often wrong when asked to give cost estimates upfront the correct solution is have a smart model (often a tech lead in human teams) do some planning and understanding, and hand over the implementation to another agent with appropriate level of intelligence and reasoning effort when the task is delegated to a less intelligent model, the smarter model also needs to continuously monitor the execution and examine outcomes to ensure things are on the right track this is a system that proved to work really well in firstmate and helped me save a lot of tokens. routing should work at the boundary between agents and subagents, not per each LLM request

pro tip - when you use OpenAI's gpt models in Codex, it uses a server-side encrypted compaction that seems to work better than anything else out there, which allows Codex to just keep hammering on long running tasks like there's infinite context window that's great, but if you run gpt in other harnesses like Pi, most of them don't inherit that by default, resulting in worse performance in long running tasks but - because of how extensible Pi is, i just found this cool extension from @alexisgallagher that enables the same server-side compaction in Pi - benchmark seems to support the argument that OpenAI server side compaction is indeed superior - so if you are using gpt models in Pi, install that extension to improve long running task performance. firstmate benefits a lot from this if you are using other harnesses, be aware of this difference and see if you can find a similar solution

alright folks i found some data - the $300/month supergrok heavy is 5x of the $30 supergrok this means with the current promotion at $100 it’s a good deal, but $300 would be paying 10x the cost for 5x quota

alright folks i found some data - the $300/month supergrok heavy is 5x of the $30 supergrok this means with the current promotion at $100 it’s a good deal, but $300 would be paying 10x the cost for 5x quota

choosing an LLM subscription these days is almost like making a hiring decision - much lower stakes obviously, but surprisingly nuanced after using many of the frontier models heavily, here’s the mental model I’ve developed: Grok 4.5 is James Bond: fast, versatile, and gets your shit done. the grok subscription also gives you image and video generation. the grok build harness gives you free access to X. the true jack of all trades GPT 5.6 Sol is Mr. Spock: the trustworthy architect and systems engineer. you don't go to it for aesthetics and personality. but it will reason through every edge case and give you the most rigorous solution. (and sometimes rewrite your entire codebase for fun) Claude Fable 5 is Winston Wolf: the legend you call to fix your hardest problems. you wish you can have the wolf around all day, but no you can't Kimi K3 is Winston Wolf’s younger brother: not quite the legend, but learned enough tricks to be a substitute GLM 5.2 is an army of Minions: affordable enough to help you move mountains through sheer volume. just be prepared to supervise some chaos

i've been enjoying the real secret weapon in Grok Build - free (seemingly unlimited?) access to all the real time conversations on X i use it to quickly understand public events and opinions, and it's completely goated - see video below. it's not sped up btw - that's jut how fast it is i'm actually surprised this isn't a built-in skill in Grok Build given this is a real differentiator, but anyway i just published how i do it as a "/whathappened" skill here - enjoy!