&. prev @PalantirTech @Umich @monitors
New York, NY
Joined August 2019
- Tweets3.2K
- Following994
- Followers1.4K
- Likes3.3K
jdrez retweeted
Who would’ve guessed sneaker inspired residential proxies would become one of the key tech unblocks for personal agents.
What’s great about @typesafeai is the plan starts you off with $5 which will last you until your grandkids graduate from college
Pacemaxxing
The real question is: what is the “right” amount of pace.
We make no mistakes? Or just enough without people dying?
Dario has written that we need to “pace the frontier,” and Sam has agreed. People may be surprised by my response: go ahead.
You guys are the frontier. By any reasonable metric — market share, revenue growth, model capability — the two of you have a duopoly on frontier intelligence. You’ve also claimed the lead is widening because of recursive self-improvement.
I don’t see what you see in the lab. If the unreleased models are scary enough that you think you should slow down, I support your decision to be responsible.
But stop pretending you need anyone else’s permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. Stop pretending METR is independent when it is intertwined with Anthropic’s investors and staff. Stop pretending you need those same evaluators to police competitors who aren’t even at the frontier.
Most of all, stop pretending the motivation to slow down is purely altruistic. You face massive product-liability exposure if your products enable a truly damaging cyberattack. The market already punishes models that behave in unpredictable or unauthorized ways. After the Hugging Face episode, it is simply good business for OpenAI and Anthropic to trade some raw power for reliability and predictability. Call it alignment if you want. It is also just giving customers what they want.
Pacing the frontier would also create breathing room for a more intelligent conversation about regulation than Bernie Sanders’ “shut it all down.” China is very unlikely to join a global agreement, as you know, and that has to be taken into account as well.
So go ahead and pace the frontier. You are the ones setting it. The easiest way not to build superintelligence is for you to agree not to build it. Demanding your preferred regulatory framework as the price of that will look like blackmail of the public and the political system. So just do it.
If you do, you’ll buy goodwill for the next conversation. If you don’t, we’ll know this was just another bid for regulatory capture — or an election-season psyop.
Instinct proves what deep information ontology can feel like
> Capturing alpha trading equities at Jane Street
> Finding the corner cases in a back office process
> Planning a trip based on my experience
The ceiling is how deep you install the sensor into the environment. Everything downstream relies on data capture positioning
YES. All of these chat + tools app do the SAME EXACT THING.
I don’t want to be told what to do
I want to offload work and know it will be done
AI SHOULD OWN THE OUTCOME
The problem with “AI teammates” (e.g. Grokbot, Claude tag, Devin) is that I actually don't want to interact with more people (real or virtual) in order to do my job better. I want the opposite actually.
I want to automate work, not have more conversations, so that I can focus, and so stuff gets done without me having to constantly guide it.
The goal of AI tooling should be automation, not interaction.
compounding is so underrated. a summer intern doing real technical work before his freshman year absolutely inflects his career path
add a little naiveté and the youngest people in the room end up setting the bar for what ambition means. (He was one of five consistently in office on weekends)
The dichotomy between the senior talent and the hungry ones just looking for reality to touch is amazing to watch
jdrez retweeted
grandmaster at work, no notes
but we've also levered the entire global economy on one company
this might be the GREATEST cross correlation risk financial markets have ever concocted
as if 2008, tech ZIRP, and purdue pharma had a child
some plausible scenarios:
1) gov manages econ figures / lab rev's to veil drawdown risk or backstop credit risk
- can someone check lev ratios of jap / life insurers in case there is no buyer of last resort on US10Y -> US30Y, otherwise will see large $$$ (monetization) of the debt
- also can / will the gov let ANT go public, probably too risky in case there's a miss
2) app layer rev's get watched like fed and one miss causes correlations to spike with 30-50% drawdowns across entire SPY
3) NVIDIA buys all competitors (Etched, etc.) / upstream suppliers (TSMC, ASML) to hedge disintermediation risk and ensure inference monopoly
3) ai app layer co's figure out secret sauce that delivers enterprise ROI's 🤫🤫
(would be largest category of all time on a rev's basis if we take the @DavidCahn6 $xyzT Q boe math @ fv)
- NVIDIA will prob fund 10-20 co's and run a CCP style Darwinian-showdown until a stable oligopoly emerges, then would be managed app layer on NVIDIA serfdom with negotiated value capture (perfect us / china mimetic rivalry achieved)
4) it all works out! no duration risk in the interim. manifest destiny at work
tradeoff's between optimization and stability in dyadic tension
@waldo47 @joeyyuwang the game is so fun, time to ride 🤝
we're now handing out open weights like party favors
every large enterprise will be looping like this soon. only a few have stopped buying and started growing:
Step 1: commoditize the model. Open weights kill businesses selling pure intelligence (at zero cost to a company that sells ads!)
Step 2: tighten your loop. turn your own workforce into RL environments
Step 3: let everyone chase weights while the intangibles compound in private
Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally. Soon we'll also release the weights for Muse Spark 1.2, our latest foundation model. Meta is a strong supporter of open source and I'm proud of these releases. Congrats to @alexandr_wang and the MSL team for all your great work on these models.
Shipping your job posting over the product itself is baller. Twenty seconds in the tui tells you more about working there than a stupid culture deck. Very well done 👏
We have had~22k sessions on `ssh superlogical.jobs` since we launched @superlogical last week. The response to it has been awesome.
This is a wish server (github.com/charmbracelet/wis…) with some extra effort for performance, rate limiting, and design.
Having seen it first hand and from talking to my peers, agents & llms are making inbound hiring a mess. Anecdotally, signal to noise on our applications was very high. Still working through them (sorry!) but responding individually to each one as the team reviews.
The ux was fun, but importantly, I think it helped us connect with people more intentionally during what is typically an impersonal process.
cost of deployment becomes identifying opportunity
cheap optimization makes accurate environments the scarcest thing in enterprise software
there are more shovels than there has ever been and still no map of where anything is buried
The future of FDE work seems closely related with all work around evals/posttraining/RL envs.
FDEs are effectively responsible for the following:
1. Define the business problem.
2. Codify the business problem into an eval rubric and environment.
3. Hillclimb the environment and output an agent/agentic workflow that solves the business problem.
Right now the process of (3) is quite manual - historically FDEs spend hundreds of hours creating bespoke software/workflows that solve the problem.
But assuming intelligence is abundant, they can effectively offload (3) to some automated optimization process. This includes RL on the model layer, and using Claude Code/Codex to optimize the harness/workflow.
Then the FDE responsibility shifts from implementing the task to defining the right goals and outcomes. In other words, they have access to /goal, and their job is more around making sure the goal, environment, and evals are correct vs. the tactical implementation details.
the only legible law firm on earth is fictional
my TL is suffering benchmark psychosis grading knowledge without reality
harvey at least had the decency to build the world first
model routing is guess optimization
Cursor built a router. Ramp built one. Stripe is paying $10B for one.
CEOs send $20M to Anthropic and see no business ROI… so the fix is a cheaper model! nobody asks if the thing is even working
every task has a best model - you just can't find it without seeing the task. routers can't see the task. they just move you sideways
The jig is up! OAI buys its 2nd consultancy this year... the next model will clear another benchmark and still won't get any real work done.
the body shop grows because intelligence was never the bottleneck - legibility is.
and the only working sensor for it is a person with a badge and a per-diem
