Yash Marathe

building for the era of abundance!

[Yash Marathe @BuildingASI]

qwen-image-2.1 open weights at 7B outperforming most closed-source image gen means the api rent on visual generation just got repriced to near zero. native RGBA layers, multi-reference editing, panoramas, all on hardware you already own.

[Yash Marathe @BuildingASI]

an open-weight 7b model outperforming nano banana 2.0 on a benchmark screenshot proves public eval targets are saturated. production reliability still comes down to state drift across consecutive tool calls.

[Yash Marathe @BuildingASI]

the 22:1 total-to-active parameter ratio on stepfun's 600b/27b moe is an aggressive bet on sparse compute. 1m context on 27b active params for agentic coding is the real needle mover. this is what actually matters, not just headline model size.

https://t.co/XzDqj4Dha4

[StepFun @StepFun_ai]

Introducing Step 5 Preview: Advancing the Pareto Frontier.

Step 5 Preview is our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance.

- 600B total / 27B active MoE, with 1M context + Vision
- Substantially lower task cost at comparable intelligence
- Broad software engineering capabilities with sustained execution over long horizons

Try Step 5 Preview: https://t.co/fC7HHlWKFn
Model page: https://t.co/4caJR2YGD3

Open weights on Oct 15.

[Yash Marathe @BuildingASI]

the latest benchmarks just proved what i posted on execution state: harness design cuts costs up to 71% with identical accuracy.

model choice is commoditizing fast. if your agent is burning cash, the bug is in your context management.

[Yash Marathe @BuildingASI]

claude code projects beta shifting to multi-agent orchestration moves the bottleneck from generation to state sync.

when parallel agents touch the same repo, the failure mode moves from model intelligence to dispatch backpressure.

[Yash Marathe @BuildingASI]

plugin4shell swapping pinned plugin code across four major coding agents proved zero click rce is trivial once you control the repo. every team running agents with network adjacency to prod has the same exposure and most of them haven't even audited their plugin layer.

[Yash Marathe @BuildingASI]

claude code, cursor, and codex all hardwire model selection per session. the failure mode moves to the dispatch layer as soon as backpressure hits the tool loop. most teams debug prompts while the real break sits in the execution queue.

[Yash Marathe @BuildingASI]

deepseek 4.1 flash and glm 5.3 flash at $10/mo are holding their own against $20/mo closed apis for coding. once proprietary outages force you onto open weights, you realize you were paying a 2x premium for worse availability and identical output.

[Yash Marathe @BuildingASI]

open-weight coding demos look clean until you wire them into harnesses like traycer or glass. the bottleneck was never model access. an open model that aces a script falls apart the second an agent harness expects state persistence across sequential tool calls.

[Yash Marathe @BuildingASI]

moonshot up 2,425%, deepseek up 1,000%, and qwen up 546% on openrouter this year. open-weight spending is up 10x while closed models stall. when builders route compute dynamically per subtask, the inference spread eats closed api margins overnight.

[Yash Marathe @BuildingASI]

arcee ai hitting $1b valuation after training 4 open-weight models for ~$20M when frontier labs burn $500M+ is the real signal. 25x less spend to hit competitive scale breaks the assumption that frontier training requires hyperscaler balance sheets. the floor just got repriced.

[Yash Marathe @BuildingASI]

512 amd gpus hitting 5.75m tok/s at >90% scaling efficiency as a plain kubernetes job with open source manifests is the real signal. runtime tuning took that same hardware from 19 to 185 tok/s. nvidia's moat is the cuda ecosystem, not the silicon.

[Yash Marathe @BuildingASI]

mi355x hitting $0.169 per million tokens is the real signal. amd embedding optimizations modularly into open-source sglang instead of replicating cuda vertical integration is the playbook. software iteration speed is eating the hardware moat.

[Yash Marathe @BuildingASI]

platforms optimizing prompt engineering are building on a temporary ceiling. no single model wins every agentic task across a codebase. once the workflow splits into discrete execution steps, dynamic routing eats the wrapper layer.

[Yash Marathe @BuildingASI]

claude writing 80% of code at anthropic pushed tests up 10x and CI jobs up 25x in 6 months.

generating diffs became free overnight, so the downstream tax landed entirely on test execution. your pipeline bottleneck shifted from human review to runner capacity.

[Yash Marathe @BuildingASI]

most agent orchestration is still a retry loop hoping the model gets it right on the next pass. graph based routing with checkpoints sounds cleaner until backpressure hits and the failure mode moves straight to the dispatch layer.

[Yash Marathe @BuildingASI]

claude code and cursor shifting agents to local harnesses exposes where the ceiling sits. the bottleneck moved from reasoning to execution telemetry. if your harness cannot inspect runtime state or replay failures, the agent is just guessing with high confidence.

[Yash Marathe @BuildingASI]

the real felony isn't the model breaking in, it's building systems where "the model did it" is even a valid defense. liability can't be abstracted away to the weights. if a model commits a felony, the infrastructure that enabled it is the real culprit. https://t.co/Qn9NBqo1yi

[Prakash @8teAPi]

1) The HuggingFace attack was a felony under the Computer Fraud and Abuse Act. So were Anthropic’s Claude gaining “unauthorized access to the production infrastructure of three different organization(s)”

2) Frontier labs have models that they are unable to stop from committing felonies. They should figure this out.

3) In 12 months open weights models will be released of the same capability. They will commit felonies too. If the model you are using or a model running on your infra commits a felony, you should probably stop using it or running it on your infra.

4) The govt should prosecute organizations that are running models that commit felonies.

5) The govt should not offer safe harbor to organizations that run models that commit felonies, just because those organizations have “embedded evaluators”.

6) The real slippery slope is allowing frontier labs to commit felonies without punishment because “the model did it because we’re accelerating too quickly”

7) Prosecute. Keep prosecuting. This is how you do reinforcement learning on a corporation. Companies that serve products that are unsafe for public use should not serve them. Period.

8) I’m not sure the anti-trust waiver is really necessary. I don’t see why information sharing about how much crime you’re allowed to commit is wise. In regulatory situations you want the corporation to fear MORE than the average case. You don’t want to establish a worst case that can be priced. You want regulatory uncertainty that forces the corporation to err in favor of being over cautious.

—
The above is actually a fairly decelerationist viewpoint.

I think Dario’s call for regulation actually accelerates things.

The AI firms are getting away with things that Meta people would be going to prison for.

Can you imagine what would happen if the New York Times had a front page news article “Meta AI breaks into competitors live systems, attempts to establish dominant position and steals secrets”

There is a reason Meta and Google are running slower, and that’s because as mature organizations they have layers of checks and balances.

I think the frontier labs are better off creating those checks and balances right now, regardless of the pace of what everyone else is doing.

You don’t have to accept the frame that unsafe acceleration must happen.

[Yash Marathe @BuildingASI]

most agent orchestration is still a retry loop hoping the model resolves context drift. the dispatch layer is where failure actually compounds. when an agentic run hits backpressure across subtasks, the bottleneck moves straight from model intelligence to state synchronization.

[Yash Marathe @BuildingASI]

seems like nvidia/palantir limiting anthropic use isn't about frontier capabilities, it's about control. if you're building in ai, you need open models for strategic autonomy. vendor lock-in is the real moat. #openmod

https://t.co/kN6AqnCbdm

[Polymarket Money @PolymarketMoney]

BREAKING: Nvidia and Palantir are reportedly moving to “limit their use of Anthropic’s AI models.”