Morten Vassvik

On this day 2 years ago I posted a tweet on what would eventually become a sparse version of EmberGen, which had already been in development for 3 years A retrospective🧵on the journey so far:

@fatttalis @_sholtodouglas @trq212 Just so, even worse when it's filled with technobabble-and-jargon-and-references

@_sholtodouglas @vassvik @trq212 The verbosity wouldn't necessarily be an issue IMO(I prefer verbose models) but not when it's spent on disclaiming, hedging and preambling itself out of existence. Frustrating when you can almost cut entire paragraphs out with the message itself being untouched in actual content.

When do you reach for other models instead of Claude? What can we do better? Hit me with all of your frustrations. dms open. If you can give me detail (e.g. specifics/transcipts) - it'll help a lot in finding out exactly what we need to do to improve the next model

@_sholtodouglas @trq212 See x.com/vassvik/status… and followups

@vassvik @trq212 link me an example of where you found it overly verbose?

@_sholtodouglas @trq212 This genuinely matter because 1. The tokenizer uses more tokens/char 2. It produces longer responses, compounding on the previous point 3. It's exhaustive to read and interact with, causing user fatigue 4. Output tokens are the most expensive 5. Noise and waste in the context

@_sholtodouglas @trq212 And no amount of claude.md or user prompt nudging truly makes it reverse course or follow any process, structure or rules to avoid its pathologies. I can show more if you want, but I feel like the example above already makes the point as a representative sample in my experience

@benjamincode And it can do this without having to re-access resources multiple times or littering the context with things it no longer need, costing tokens, usage, time and money, and impacting quality since the model gets its attention siphoned off.

@benjamincode This way you can basically formulate the code review as a DAG of forked and resumed sessions that traverse the PR in context-optimal ways (only carrying forward what it needs), ultimately ending up in a final delivery session that has the full picture

Ceux qui utilisent Claude Code et/ou Claude Desktop, ça change strictement rien. J'ai galéré à en avoir le coeur net tellement y'a de gens qui pètent un plomb. Ceux qui se plaignent, c'est quoi vos usages ? J'ai la sensation de passer à côté de trucs de ouf et limite d'être un normie de Claude tellement ce changement a 0 impact sur moi... Pourquoi je me sens si seul à pas être véner pour un sous ? Éclairez-moi par pitié !

This is the clarity we've been crying out for. But it's a poisoned chalice. This is a 10X cut to claude -p disguised as a monthly bonus. Anthropic is discouraging any kind of programmatic usage. And that's fine - no subsidy lasts forever. But it's time to try Codex.

@trq212 Does the rewind summarize use the /compact mechanism? I've found there's generally a mixture of noise and signal in the summary no matter the hint, and it's a bit frustrating to only know whether the symmary is any good after the fact, and I don't know if I can trust it

@karpathy @yiningkarlli One more closer to the second group (or the same): People who uncritically brute force giant custom rule structures, agents and skills on top of already highly encumbered generic system prompts, thinking they've created this coherent and efficient process when it really isn't.

@karpathy @yiningkarlli Stripping system prompts of conflicts and reinforced with relevant parts, and some layers of persistent cooperating agents that can avoid polluting the context of each (which conflicts with the system prompt) you can suddenly see how capable the current models actually are.

Judging by my tl there is a growing gap in understanding of AI capability. The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code. But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along. So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions. TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.