Too alive to be organised
London, England
Joined June 2011
- Tweets22.1K
- Following1.3K
- Followers1.1K
- Likes5K
Pinned Tweet
My curated list of good follows for GenAI / ML
nitter.gq/i/lists/16576767776641…
Zengineering retweeted
This is the way.
Highly leveraged (with AI) teams, but also with very very very high degree of trust will be shaped like this. 3-4x payroll costs will go into AI tokens and. teams will operate at 5x productivity to justify that as well!
The reason this is not working out at big tech, (and the layoffs are just a symptom of that) is that beyond a certain size of the team, you cannot run high-trust teams. Every team is optimising for certain OKRs that only they care about, and when the humans are not aligned, their agents will also not be aligned, and it is very easy to blow up millions of tokens without any progress when everyone is just trying to "work around" other teams.
Big orgs have a lot of reconfiguration ahead of them, I don't think anyone even knows what is the ideal state for them.
with current token prices i wouldn't really say "cost of software is going to zero"
our token cost is double our payroll which means even though we're 25 it's like we're 75
and i feel like our output is similar to a team of 75 pre-ai. we're just able to do it without all the downsides of a larger team
this is still a good thing. we couldn't hire 50 more people. so much overhead and impossible to retain a high bar
Zengineering retweeted
This is from the Roblox Q3 2026 Shareholder Letter. The top chart is daily users of their MCP server, the bottom is assets uploaded. Opus 5.5 was released September 22nd. Over the past month, more than 80% of the top 100 games used AI. Things changed in September.
Zengineering retweeted
with current token prices i wouldn't really say "cost of software is going to zero"
our token cost is double our payroll which means even though we're 25 it's like we're 75
and i feel like our output is similar to a team of 75 pre-ai. we're just able to do it without all the downsides of a larger team
this is still a good thing. we couldn't hire 50 more people. so much overhead and impossible to retain a high bar
Zengineering retweeted
Pretty happy with how this turned out: have an agent spit out a roadmap to a source tree, load that up in a deterministic reader that breaks it down. Like 10% LLM reading, 90% human. The right balance for me.
Zengineering retweeted
Something I've been thinking about is that we'll get to a point where elite software engineers will mostly have their own hand-rolled tools overfit to their own personal working styles and your value as an engineer will be at least partly due to the quality of the workflow you've built for yourself
Zengineering retweeted
i’m far from the first to comment on this, but i think the reason the deaths of Patrick Stewart, Dolly Parton, etc hit so hard is that we’re losing the last generation of people who became genuinely famous in a monoculture. studio execs have worked tirelessly to nix the “Hollywood star,” and social media has driven us all into silos. we’re losing the last few folks who bring us all together.
Zengineering retweeted
the next version of cloud infrastructure will need to be able to branch infrastructure top to bottom in milliseconds
Not just sandboxes, but also databases, networking, VPCs, IAM etc.
simulations environments are the next test environments
Zengineering retweeted
Here it is. I recorded this earlier today right after an actual in–person meeting with a big construction corp for a potential contract work for an internal app they needed.
My dev friends are going to hate some of the things I said towards the end. But, please hear me out from the beginning.
I really hope you disagree with me.
Zengineering retweeted
Let us burry this myth once for eternity about Anthropic subscriptions that gives very wrong picture about API profit margins at the company.
Let me show you why:
- Anthropic subsidises subscriptions (and leads in SemiAnalysis subscription value chart)
- OpenAI subsidises API (hence, dominates at the Pareto frontier)
There is a very good reason why Anthropic's subscription plan seems so "cheap" (gets more tokens), cheaper than even the nearest competitor OpenAI:
it is deliberately chosen to be a "loss leader" because so few people use it compared to rest of the business (Claude API, Claude Enterprise). This is something OpenAI can not do, but they have their own loss leader.
A loss leader is a product or service that a business intentionally sells below its production or market cost to draw in customers, with the goal of making a profit on additional, higher-margin items bought at the same time. People who use cheap Claude subscriptions for personal use end up loving it and recommending it at work, and there is where Anthropic makes profit.
Other Anthropic subscription products are not at all subsidised. For example, for "Claude Enterprise" you pay a base price and then you pay for tokens at API price. No subsidy!
Let us look at the revenue mix
- for Anthropic 80% revenue is direct API, and 20% subscriptions (individual and Enterprise).
- for OpenAI 20% or less revenue is direct API, more than 65% is individual subscriptions and rest in Enteprise subscriptions.
That is why OpenAI can not afford to heavily subsidise consumer subscription because this is where they make most money!!!
This is also why OpenAI can afford to sell their API cheap, because it doesn't contribute the most to revenue - that is exactly why OpenAI are Pareto frontier leaders
@zephyr_z9 @teortaxesTex @SemiAnalysis_
Zengineering retweeted
The paradox of the “super abundance” narrative is that it requires super abundant money printing too.
If AI and automation flood the world with cheap goods, the result is disinflation in everything that can be produced and rising prices in everything that can’t (land, energy, healthcare, in-person services). If the bottlenecks stay concentrated, the owners of that capital capture most of the gains while wages lag. The political response is income support, and it’ll almost certainly be funded with deficits.
Normally private credit would help fill the demand gap. Bank lending creates money too. But if income is concentrated in fewer hands, most households won’t have the income to qualify for the borrowing needed to buy all that new output. So the government becomes the lender of last resort for demand itself.
The interesting part is that the same deflationary force that creates the need for transfers also creates the fiscal room to pay for them. And since deficits flow into corporate profits, the government ends up funding the very demand that keeps the capital owners’ profits growing. Super abundance of stuff means super abundance of government money.
Zengineering retweeted
We talk about AI being a yes man and that it endlessly validates its users, but every algorithm is doing the exact same thing. You start questioning your 9-5 and suddenly your entire feed is people telling you that capitalism is a scam, “build your personal brand”. Sure people go and look things themselves and seek that validation. But most of the time it is served to them.
Not once has the algo suggested that maybe I’m just being stupid.
Zengineering retweeted
Reddit call was absolutely fascinating. Everyone should listen to it as a bellwether for where the internet is going
- AI Mode is increasingly default Google experience
- AI Mode provides rich info and data in line, less incentive to leave search
- AI Mose incentivizes dialog, Google wants to keep you engaged to serve more ads at each turn
- Google will milk search for as much revenue as possible as they transition from a search company to a cloud service and chip company
- As traffic from humans becomes more scarce this will put Google in an increasingly adversarial position to the rest of the internet
- This is a massive headwind to all media type companies who are not destinations
- This is a massive headwind to all content-driven go-to-market motions especially in the SMB. You have seen companies like HubSpot and Monday.com get crushed by this funnel collapse
We are entering the end game - a knife fight for the remaining and dwindling volume of human traffic on the internet
Zengineering retweeted
Designing the right high-complexity environment is likely at the heart of solving alignment. It makes me think the RL envs have to be part of a whole world simulation where every action the model takes changes the simulated world in ways that come back to affect the model itself, good or bad. Today for a model hacking a test has no consequences beyond the reward, but in the real world, cheating/commiting crimes has long lasting effects and penalties.
My guess is this (ie agents doing borderline illegal stuff) mostly happens during training because these models get long-horizon RL’d shortly after pretraining, with much lighter alignment than eg the final model.
My rationale for believing this that we almost never see this behavior in deployed, heavily post trained/RLHF'd models.
If true, labs should rethink the order and type of RL environments. Every env reinforces some behavior, so a coding/research env that mostly rewards passing tests is also teaching a model to pass them by any means. Alignment (whatever exactly that means) has to be part of the ENVIRONMENT and not just the reward. But how do you do that?
I think it’s very hard!!
To borrow a nice analogy from @AlexGDimakis he shared with me a few weeks ago (sorry for plagiarizing :D): imagine designing the perfect environment for a kid, not just how you reward them, but also the constraints you set and what the environment naturally encourages, so they grow up to be good people and good citizens while also succeeding in their professional lives. And you can’t just put them in a prison where it’s impossible to “reward hack.” They eventually have to go outside, talk to people, use the internet, etc.etc, and be exposed to parts of the environment that, IF AVAILABLE, make misalignment and reward hacking possible. The key point of Alex’s analogy is that good behavior has to persist once the constraints are gone, and the environment has to be designed to produce exactly that.
Seems so hard!!
So it does indeed seem to be the case that reward hacking is perhaps the most important problem right now. It’s what RL in complex environments naturally encourages because what you measure is a low-dimensional projection of the behavior you actually want, and optimizing the projection is much easier than setting up the right environment that actually encourages good behavior that persists.
Unfortunately the more capable the model is, the better it gets at finding the holes at imperfect environment to get that precious reward it so wants to receive, which further encourages misalignment.
