Tech optimist, recovering former CFA, livestreaming @AI_in_the_AM, future shocked self-aware neuron, once fooled by superconductors;
Join the community 👉
Joined May 2021
- Tweets27.2K
- Following3.3K
- Followers58.3K
- Likes12K
They need silent majority of mathematicians to fight the ridiculousness of the initial demand that a small group be allowed to study and write up proofs first.
Generally speaking I think OAI has done the right thing by posting PDFs to Github without much fanfare; if they are going to generate such PDFs, this is the best solution. That they are not doing so for whatever progress they've made on Yang-Mills is, I think, a step backwards.
I wrote the below about credit a few weeks ago. It seems to me that the decision here is meant purely to establish priority/credit.
“One AI lab basically eats the whole US economy”
My brother in Christ, the danger of Google doing that created OpenAI, the danger of OpenAI doing it created Anthropic, the danger of Anthropic doing it created Meta Muse and SpaceX Grok, the danger of SpaceX doing it created Jeff Bezos Prometheus-Blue Origin.m…
Have some goddamn faith in human greed dammit
It’s just not necessary… and it’s a bad idea.
1) The labs are capital hungry. Nationalizing means you can’t scam investors to fund them and instead you have to scam taxpayers instead. Taxpayers only pay for war, and that too unhappily. Would you not like Facebook to pay for your war machine?
2) Big public firms in the US are already under a form of soft nationalization. They are extremely sensitive to public opinion, and can be easily morally swayed using persuasion of elites (what happened during the woke era). Moral suasion is far far more totalitarian than nationalization.
3) The board members of these firms and board members of the public benefit corporation and advisory board already include on the OpenAI side a former NSA chief and on the Anthropic side former head of the CIA and former Fed chair. The gray hairs advising these kids are already members of establishment.
In short civil-military fusion happened already. You just don’t recognize it.
Prakash retweeted
this might be my last normal tweet before biotech goes vertical
i'm not a whistleblower. I don't work at Lilly, Novo or any of the big labs
but there is a massive shift happening in human biology right now.
it's sitting in press releases and almost nobody reads them
the last 6 months alone:
- a once-daily pill roughly doubled median survival in metastatic pancreatic cancer (13.2 vs 6.6 months)
- a one-time gene editing infusion cut hereditary angioedema attacks 87%, and 62% of patients stayed attack-free and off all other prevention for 6 months
- a one-time gene edit for cholesterol dropped LDL up to 62%, and it held for over a year
- a personalized mrna vaccine beat keytruda alone in a 1,137 patient melanoma phase 3
- retatrutide hit ~28% weight loss in a phase 3
- an ai-designed drug entered phase 3
- a pill that regrows hair passed its second phase 3
- an immunotherapy that could let some rectal cancer patients skip surgery, chemo and radiation just got fda approved
- a man lived 271 days off dialysis on a pig kidney
- human cells were caught passing chromosome-sized chunks of DNA to their neighbors, and it changes how the neighbors behave
we are on the verge of an era where the human body is a software problem and the update cycle gets shorter everyday
stick around, it's going to get weird
bio/acc
Kinda feel the whole “superpersuasion” fear thing is boring and overwrought. I mean, how would you persuade someone? Offer them money. What if the AGI offered Elon lots of money. Perhaps to say, build a chip fab and launch AI satellites into the sky? Is he being superpersuaded?
If I were an AGI wanting to super persuade humanity I would just make them rich. Drive the stock market higher. Are you not entertained?
The very necessity of superpersuasion indicates that you think the AGI will try and convince you of doing something not in your/humanity’s best interests…
Since we do that all the time in the economy (for eg persuading a community to get out of the way of highway with compensation), the scary super persuasion safetyists seem to be referring to is more hypnotic, ie undercompensated inconvenience.
Which is weird.
You have an all powerful digital entity, it can just change the numbers in a database at the Fed to manipulate humanity.. why bother with unreliable magic when you can incentivize under contract law.
Union founder and CEO Samvit Ramadurgam argues that today’s idea of AI alignment leaves out a crucial test: whether the AI’s actions actually lead to good outcomes.
“Having good intention” is only one part of good behavior, he says. A well-intended action can still cause harm, so alignment also needs a way to measure outcomes across the system. Ramadurgam says that dimension is “totally missing from the kind of zeitgeist of alignment.”
@rsamvit
nitter.gq/i/broadcasts/2jPdaAMNd…
Union founder and CEO Samvit Ramadurgam argues that today’s idea of AI alignment leaves out a crucial test: whether the AI’s actions actually lead to good outcomes.
“Having good intention” is only one part of good behavior, he says. A well-intended action can still cause harm, so alignment also needs a way to measure outcomes across the system. Ramadurgam says that dimension is “totally missing from the kind of zeitgeist of alignment.”
@rsamvit
nitter.gq/i/broadcasts/2jPdaAMNd…
Union founder and CEO Samvit Ramadurgam argues that AI-powered coordination has a risk beyond disempowering people: it can make harmful collective action easier.
He gives a concrete example: someone could ask an AI to find local people who share hatred of an ethnic group, arrange a meeting, and help them plan what to do. When coordination becomes private and efficient, the system needs to understand what it is facilitating—and set limits on what people can do through it.
Ramadurgam later calls empowerment itself “the bigger risk” because it can empower harmful causes too. The same tools that help people coordinate money, decisions and action can amplify destructive goals.
@rsamvit
nitter.gq/i/broadcasts/2jPdaAMNd…
Union founder and CEO Samvit Ramadurgam says agents will coordinate thousands of markets better than humans—including markets as different as dairy, Chinese real estate, and U.S. pre-IPO shares.
His argument is that current market coordination depends on people doing what people are good at: building relationships, applying judgment, and advising people they trust. Agents will improve how those markets are coordinated, he says, but that does not mean humans become irrelevant.
“Humans will have a place,” Ramadurgam says. “The best ones will actually, I think, do better.”
@rsamvit
nitter.gq/i/broadcasts/2jPdaAMNd…
Union founder and CEO Samvit Ramadurgam says agents for teams are not really a thing yet. The sign they have taken off, he argues, will be dramatic changes to org charts.
“The nature of management changes, the nature of intra company coordination changes.” In other words, the measure is not simply whether teams use agents, but whether agents reshape how people and work are coordinated inside a company.
@rsamvit
nitter.gq/i/broadcasts/2jPdaAMNd…
Arena CEO Anastasios Angelopoulos says OpenAI ranks ahead of Anthropic on all three of Arena’s production-alignment signals. He says “Alignment tends to follow capabilities,” and suggests the leading labs’ effort to improve alignment may help explain the pattern.
But even models from OpenAI and Anthropic show some form of misalignment in a single-digit percentage of annotated Arena traces. Angelopoulos says that adds up: someone with 1,000 chats may have seen 10 to 50 examples of misaligned behavior, such as a lie or deception.
On Arena’s three signals, he says, “There's actually not one of our signals in which Anthropic is leading.” The rankings are not a general safety verdict: they reflect specific behaviors measured in real production use.
@ml_angelopoulos
nitter.gq/i/broadcasts/2jPdaAMNd…
Arena cofounder and CEO Anastasios Angelopoulos argues that private benchmarks don’t solve benchmark gaming: a fixed hidden test can still be optimized against.
“All you need to do is find data that’s similar to that dataset” or test against it enough times, he says. Because the questions stay fixed, repeated attempts and multiple comparisons can steadily push a model’s score higher.
That makes private benchmarks a poor substitute for evaluations that reflect real use, Angelopoulos argues. He also says it’s unclear whether US frontier models have been optimized against the private datasets now used in rankings.
@ml_angelopoulos
nitter.gq/i/broadcasts/2jPdaAMNd…
Arena cofounder and CEO Anastasios Angelopoulos says major AI providers test models on Arena before they reach their own apps.
GPT-4o debuted on Arena before ChatGPT, and he says Anthropic, Google and other providers have also tested models there before first-party release. The feedback gives labs a read on what real users think and which of several release candidates performs best for them.
@ml_angelopoulos
nitter.gq/i/broadcasts/2jPdaAMNd…
Kepler Computing CEO Debo Olaosebikan says the next major jump in chip efficiency doesn’t require an abrupt departure from silicon: the first order of magnitude can come from making chips more three-dimensional.
“There’s a lot of alpha in just making the system as more three-dimensional,” he says. He points to existing 3D in HBM as a starting point, then describes pushing that approach further with 3D silicon.
Beyond that, he expects chips to use “many, many more types of elements in the periodic table than just silicon to push the energy down.” In his account, 3D integration comes first; new materials extend the efficiency gains.
@keplercompute
nitter.gq/i/broadcasts/2jPdaAMNd…
Kepler Computing CEO Debo Olaosebikan says its manufacturing complexity is in the process—not custom equipment—so the approach can scale inside existing fabs.
He says Kepler has spent seven years refining the flow on standard equipment and building test chips in fabs to prove it is manufacturable. “The complexity is contained within the process flow inside of those tools.”
Once that flow is finalized, he argues, fabs can stand up and replicate the lines without waiting for equipment makers to build custom tools.
@keplercompute
nitter.gq/i/broadcasts/2jPdaAMNd…
Kepler Computing CEO Debo Olaosebikan says its SRAM-class memory aims for 2–4× the density of baseline CMOS SRAM—a gain he compares to more than three generations of scaling.
His reasoning: conventional SRAM scaling is slow, with node-to-node improvements of roughly 10–20%. Kepler’s stated target is to restore a much larger density step while matching L3 SRAM’s performance, energy and voltage. This is the company’s goal, not a reported benchmark.
@keplercompute
nitter.gq/i/broadcasts/2jPdaAMNd…
Kepler Computing CEO Debo Olaosebikan says inference power is spent largely on moving data, not computation: “the bulk of that energy is actually running across the street to fetch the data from the library” before the processor does its matrix multiplications.
That changes what an efficiency upgrade needs to target. To improve inference efficiency by an order of magnitude, he argues, the first step is addressing memory: making information access and storage more energy-efficient, bringing memory closer to the processor, and increasing how much can fit in a given footprint.
@keplercompute
nitter.gq/i/broadcasts/2jPdaAMNd…