Founder & CTO @khalani_network | post grad philosophy of mind at GCAS Dublin prev SC & formal verification @MakerDAO 💲prev eng @NervosNetwork
In a DB
Joined April 2018
- Tweets3.9K
- Following2.6K
- Followers2.1K
- Likes1.6K
I wonder how many people had to watch and sign off on this video.
Every character is loathsome, depicting rather accurately an archetype that is increasingly common in our society.
Was the goal to repulse? Or have the makers fallen so far they were unaware that the family portrayed here is weak; is void of mutual respect; is starved for values.
The father is the ultimate antagonist, but so pitiful that it takes a moment to recognize that.
The dog is, ultimately, the protagonist.
Tannr 🦾🤖 retweeted
We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case).
I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves.
We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them:
"/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started"
I imagine we'll have some backend services that could be replaced entirely by this new architecture.
Tannr 🦾🤖 retweeted
A problem in probabilistic reasoning almost everybody gets wrong. It’s astounding how unintuitive it is for most people (myself included!) to reason probabilistically. Might bring this up with in my Critical Thinking class later in the semester.
Tannr 🦾🤖 retweeted
I don't know who needs to hear this, but a degree in philosophy is one of the best preparations for law school
And thus we observe the return to verification-centric architectures (realized and idealized by UTXO chains like @NervosNetwork and Bitcoin).
Whether by means of cryptography or other forms of succinct verification, the concept is the same: don’t replicate globally what need not be replicated globally.
The cryptographic world computer:
vitalik.eth.limo/general/202…
My attempt to express in somewhat concise terms the true meaning of basically everything planned to happen to Ethereum starting from the fork after Hegota. It's really not just a blockchain anymore. It's a hybrid architecture that combines together blockchains and modern cryptography, to enable much more powerful properties.
FOCIL, EIP-8288, Lean consensus, state management, formal verification, advanced mempool improvements (including privacy), and the longer-term specter of obfuscation all mentioned.
Oh wait, I just realized you're arguing *against* targeting Hegel due to the fact that Hegel "led to" Marx.
That's different. If Marx's encounter with Hegel's works was material in the development of Marxist Dialectics, then in that direction, there is a critical link.
Marx isn't a good representation of Hegel, though. Marx is but a footnote to Hegel, one could say. And an embarrassing one, at that.
The teachings of Christ led to the inquisition by the Roman Catholic Church.
The violence involved in carrying out that mission probably violates the spirit of Christ's teachings.
A person's actions can be inspired by the words or actions of another. It is not the latter's fault if the former lacks the intelligence to understand those so-called "inspiring" words or actions.
The problem is Marx. Marx is accountable to his own ideas. That he looked at the words written by Hegel and then applied his (inadequate) understanding is not Hegel's fault, lol.
Tannr 🦾🤖 retweeted
William James: 'The worst of criticizing Hegel is that the very arguments we use against him give forth strange and hollow sounds that make them seem almost as fantastic as the errors to which they are addressed. The sense of a universal mirage, of a ghostly unreality, steals over us, which is the very moonlit atmosphere of Hegelism itself. What wonder then if, instead of converting, our words do but rejoice and delight those already baptized in the faith of confusion? To their charmed senses we all seem children of Hegel together, only some of us have not the wit to know our own father.'
(The Will to Believe)
It is remarks like these that demonstrate how misunderstood and poorly taught Hegel is in today’s classrooms
I would love to be a plane ticket.
What it’s like to be a plane ticket is an even more intriguing question than what it’s like to be a bat.
How much $$ to be a plane ticket?
Serafim is doing really cool work at the intersection of LLMs and various forms of formal and semi formal reasoning.
I evaluated the latest models, GPT-6 Sol, Luna, and Claude Opus 5.5 on induction (post from yesterday).
GPT-6 Sol is doing very well, somewhat behind Opus 5.5 but at 1/2 the price. GPT-6 Luna is doing slightly better than GPT-5.6 Luna. 14 responses are still running on high thinking effort after failing on max and xhigh. I don't expect them to significantly raise the rate of success.
Overall, I think this is a success for the new GPT-6 models. The new Sol is significantly stronger than the old Sol at 2/3 the price, and slightly behind Opus 5.5. Luna is slightly above the previous Luna at 1/3 the price!
So, both new GPT-6 models are moving solidly in the right direction of intelligence-per-buck.
Github: github.com/SerafimBatzoglou/…
ICML 2026 paper: arxiv.org/abs/2602.18956
“GPT-6 Sol is doing very well, somewhat behind Opus 5.5 but at 1/2 the price.”
This makes me wonder: is there a metric you’re thinking of that could adequately represent inductive-power-per-dollar? And how would the fact that they have different ceilings affect that metric? If. model A is half the price of model B per inductive “unit”, it’s more cost effective. But if model B’s ceiling is higher, then I suppose model A, past its ceiling, would discretely change…
Confidentiality is better than publicity for those who need informational discretion.
Privacy is better than confidentiality for those who need informational isolation that doesn’t depend on the discretion of another.
Not criticizing here, but curiously pushing back a little bit: the thread starts with claims of privacy, and then notes that the trusted operator has a global view of the zone, thereby ensuring transactions are confidential.
That’s arguably an equivocation. It is easy to make honestly if one is not using precise definitions of “private” and “confidential”, but generally they mean different things.
“Confidential” seems to me the more accurate term here, as I would expect “private” to imply that the details about the transactions are operator-blind as well.
Broadly, if the degree to which the information in a system is isolated from others is at all a function of the decisions of a rational/normative agent, then the system is not private, imo.
If I’ve misread things I’m happy to be corrected. I’m going off of the thread only — will dive into the technical report as well, of course.