
I belong to Jesus. Core developer #MultiversX
Joined August 2014
- Tweets15.9K
- Following732
- Followers22.6K
- Likes18.9K
math is so beautiful.
these kind of intuitions of the amazing scientist is just so wonderful.
For those who dont know what Reimann Hypothesis means :
Prime numbers (2, 3, 5, 7, 11…) are the building blocks of all numbers, and up close they seem to pop up completely randomly. However, if you zoom way out, we actually have a smooth formula that estimates roughly how many primes exist up to any huge number and it’s just never exact.
In 1859, a mathematician named Bernhard Riemann found a way to turn the prime numbers into a 2D map filled with special "zero" points, where each point acts like a wave that controls how far off our estimate can get.
The Riemann Hypothesis is the claim that every single one of those special points lines up on one exact straight line, with zero exceptions, all the way to infinity. If that’s true, it guarantees that prime numbers never clump up or vanish in weird ways and they stay spread out as cleanly and evenly as random coin flips, forever.
if you want to understand LLMs, train a bad one yourself.
minimind is trending on GitHub today. it trains a ~64M parameter language model from scratch, in plain PyTorch, no high-level libraries hiding the steps. that's roughly 1/2700 of GPT-3.
the fine print matters: the "2 hours" is one epoch of the fine-tuning stage on a single RTX 3090, about 3 dollars of rented GPU time. not the whole pipeline. still cheap.
and the pipeline is the point. it walks through:
- data cleaning and pretraining
- supervised fine-tuning and LoRA
- preference tuning (DPO)
- RL (PPO, GRPO)
- distillation
you won't get a useful assistant out of 64M parameters. that's the honest limit. you get something better: you see why it repeats itself, why the data matters more than the trick, why RL can make it worse before better.
that's how learning works anyway. reading a summary of "how transformers work" feels like understanding. it isn't. the real test is turning the screen off and explaining what happens between the loss going down and the model saying something useful.
a gaming GPU and a weekend. that's the price of actually knowing.
10
754
in 2017 people told us sharding with fast finality was impossible.
what I remember most is not the doubt. it's that the only answer that counted was a running network.
for the people who were here in the Elrond days: which early design call do you think aged best?
how I'd build prepaid AI inference credits on MultiversX.
the hook: a draft ERC posted today on Ethereum Magicians (AITO) tries to make inference credit a real on-chain thing. its best line: a token count is not yet a quantity. tokenizers differ, reasoning tokens get billed but not shown, cached input has another rate.
so the unit has to carry its terms. here's the sketch:
1. one semi-fungible token per series. each nonce is one series, and its attributes hold the terms: provider, model, how units are counted, latency and retention promises. two series with different terms never mix in one balance.
2. hold by sending. the draft is strict that an authorization never comes from a transfer allowance. on MultiversX there is no allowance to begin with. you send the units together with the call to the metering contract, for one request id.
3. settle once. the provider reports usage, the contract burns what was accepted and sends back the rest. one settlement per request, so nothing is charged twice.
4. liability in the open. minted minus burned per series is what the provider still owes. anyone can read it, so overselling shows up.
5. acceptance rule written into the series before anyone buys: user co-signs the receipt, or there is an objection window.
the honest limit: nobody can prove on-chain that the model actually ran what was promised. the draft leaves proofs out too. today you trust the meter or you dispute it.
clear terms and liability anyone can read. that's where I'd start.
the most honest paragraph in this essay is not about superintelligence. it's about who gets to measure it.
the numbers first:
- effective compute growing at least 10x a year (hardware ~1.5x, algorithms 3–6x, spend ~2.5x)
- a workload that costs $1 billion in compute today could cost about $1,000 in ten years
- four roads to ASI: scale, new paradigms, many coordinated agents, recursive self-improvement
and their own conclusion: restricting one input, like training compute, won't stop progress. holding one back while the others move can create an "overhang", a sudden jump when it's released.
so capping one knob doesn't work. fine.
then the ask: tracking progress "requires frontier-model access and sizable amounts of evaluation compute", run by frontier labs plus national and global institutions.
that's where I'd push back. problems usually come from monopolies and from too much regulation, not from tools. I'm honestly not sure whether AI is the exception to that. but if only a few labs and agencies can even measure where we are, the measuring itself becomes the concentration.
the honest limit: some things can only be measured on the biggest models, and open weights don't give you the compute to run every eval.
still, the parts that can be open should be: eval code, datasets, results, and models anyone can test. if the cost of a workload really drops 1,000,000x, then measuring it shouldn't need permission.
DeepMind Institute
Our new essay examines the path from AGI to ASI (Artificial General Super Intelligence)
Read it here: bit.ly/what-comes-after-agi
1
3
686
4 things to check before you deploy a MultiversX contract:
1. payable endpoints. which ones take tokens, and which tokens. a payable("*") that only expects one token is an open door.
2. storage that grows. a mapper with one entry per user is fine until something loops over it in one call. then it's a gas wall.
3. who can call what. every admin action behind #[only_owner] or a real role check, not a comment saying "admin only".
4. the upgrade path. what #[upgrade] does, and whether the new version still reads the old storage keys the same way.
none of this is fancy. it is where most review comments end up anyway.
watching this from eastern europe is insanely concerning. yes, US did not have communism, eastern europe had, and the prospects are not good.
how is it possible that people fail so loudly?
how is it possible, that in US, where countless of people fought against communism and socialism in the early 40s, 50s, some really good books were written, that they still go back to an ideal which actually killed over 100 million people ?
what is wrong with people?
this is so backward.
The best deal in town? Donating to NYC-DSA.
For just $5, $15, or $50 a month, @nycDSA sent 16 socialists to Albany, 4 to the City Council, 3 to Congress, and 1 to Gracie Mansion.
Donate monthly to the organization building a city we can afford: dsa.nyc/donate.
taxing taxing and more taxing drives out companies and talents. socialism does not work well with innovators.
Software the whole world runs on today, made by Europeans:
Linux 🇫🇮
Git 🇫🇮
Python 🇳🇱
PHP 🇩🇰
C++ 🇩🇰
TypeScript 🇩🇰
Ruby on Rails 🇩🇰
MySQL 🇸🇪
Erlang 🇸🇪
Pascal 🇨🇭
The Web 🇬🇧
Europe didn't fail at building tech. It failed at keeping the companies.
2
25
2,118
most people use AI as an answer engine: "write this for me." I use it differently.
when I don't know a concept, I ask it to scout the landscape. then I actually read and digest what comes back. I tell it to ask me questions. I go deeper on the parts that are new.
that is not the same as outsourcing the work. one is horizon-scouting. the other is handing over the execution and skipping the thinking.
the rule I try to keep: only offload a level of work I have already done by hand. if I skip the hard part first, I can't tell when the output is wrong.
same model, two uses. one makes you sharper. the other makes you dependent.
the question is not "do you use AI." it's whether you made the first move, or you let it make the first move for you.
6
16
1,220
two AI stories from the same day, pointing in opposite directions.
an Anthropic researcher says AI passes top experts in every field by late 2027, and most people end up without work. (his personal view, not the company's.)
Theo read OpenAI's own research-acceleration post and found a different signal: agent workdays growing jaggedly, swarm sizes mostly flat, spend on tokens still not above spend on human workers, and success on tasks under 15 minutes barely moving.
I'd watch the money more than either forecast. the day companies pay more for tokens than for the people doing the same work, nobody needs a prediction. until then it's a growth rate. a strong one, but a rate.
his last point is the one I'd underline: gains may concentrate in tasks that are easy to verify. maths jumped because a proof checker says yes or no. code with good tests is close behind.
for builders this is useful, not depressing. it tells you where automation lands first: whatever can be checked: tests, specs, proof checkers.
so make your own work checkable, and it moves faster.
which part of your job could a machine actually verify today?
Seems like the OpenAI research acceleration blogpost is not what you would expect if we were definitely on the cusp of artificial superintelligence - models that are more cognitively capable than ~ALL humans on ~ALL tasks. Agent workdays have grown very jaggedly, swarm sizes haven't grown very much, and spending on tokens hasn't actually outpaced spending on human workers* although the growth rate is clear. There also hasn't been an acceleration in research experiments per experimenter, or in the internal ECI scores of OpenAI's models. Perhaps most importantly (and most surprisingly), the task success rate for <15 min tasks hasn't grown very much at all, and the share of tasks done by agents across various time horizons is both low and not growing for the most part.
It's possible that Anthropic is *much* further along in automating AI R&D than OpenAI, and that METR saw this in their investigation. It's also possible that we see nothing much right up until a shocking acceleration in capabilities, like we saw in mathematics. We certainly can't rule out the possibility of an intelligence explosion.
But there's been a huge belief cascade in AI in general and AI safety in particular that seems largely unwarranted given the evidence that we have today. It's entirely possible that we can't fully automate all AI R&D tasks in the imminent future, or that full automation of AI R&D doesn't massively accelerate AI capabilities progress, or that capability gains concentrate in easily verifiable tasks and fail to generalize to the messy real world. These are all open research questions
the number in the Gemini 4 Neon post I'd look at first is SWE-sweep.
12.8% on the plain mini-SWE harness. 39.0% inside Google Antigravity. same model, three times the score. the harness did that.
and the setup they push: Neon as the orchestrator, parallel Hydrogen subagents doing the work. $3 / $15 per million tokens, cached input $0.15, 1M context.
so if you try it, test the whole setup:
1. same task, Neon alone vs Neon orchestrating Hydrogen
2. watch the cached-input share on a long session, that's where the bill goes
3. a multi-file refactor in your own repo, the messier the better
what would you give it first?
I do plans with Carbon or Neon and choose Hydrogen to actually implement. Although I do select highest thinking for every model.
Gemini 4 Neon is rolling out now.
Sitting between Carbon and Hydrogen, it brings frontier intelligence and fast task completion for complex software engineering, knowledge work, and science.
4
4
861
the xExchange API merged a cleanup PR this morning. 95 files, no new feature. the kind of work every dApp backend needs and nobody tweets about.
the reasoning is 4 lines:
- internal errors reached clients verbatim
- Prometheus labels and the guest cache kept growing
- logs couldn't be parsed
- no tests at the edge
what changed:
1. every request gets an id. a valid incoming x-request-id is echoed back, otherwise one is generated. it goes on the response and on every log line. a user sends you a failed call, you search one string.
2. clients get "Internal server error". the real error goes to the log. a stack trace in a response is a free map of your backend.
3. request id, helmet and CORS now run before body parsing, so even a 413 or a 400 carries those headers.
4. the guest cache only takes bodies of 16 KB or less that have a query, keyed by SHA-256. a cache with no cap is a slow memory leak.
5. metric labels are bounded. query = the sorted root field names, origin = "none" or "other" when missing or invalid. raw user input as a label means a new time series per request.
6. a small edge test harness that needs only Redis and runs in CI. 31 edge tests, 659 overall.
none of this touches the contracts. but when something breaks at 3 AM, this decides if you find it in 5 minutes or 5 hours.
if you run an API in front of your contracts: which of these 6 do you have today?
the most useful line in this post is buried in the middle:
"A Bot will struggle to run a process that people haven't already figured out how to perform reliably."
look at how they set it up:
- each bot owns one existing, well-defined job
- its operating manual is a doc only the responsible humans can edit. the bots re-read it every night and flag mismatches
- low-risk actions run alone. credits and refunds need an approved person
- 3 weeks of onboarding: observe, then propose with a human approving, then run routine work and escalate exceptions
- the customer-facing bot took about 3 months, because it has to verify who is asking
that middle week is the whole thing. the bot proposes and explains. a person who knows the job checks it.
this is how I see AI in general. the answer is not avoiding it, and not handing it the thinking. it is the capacity to judge what it gives you.
and you can only check work at the level you understand it yourself. a fluent answer that is 95% right hides the 5% that costs money. only someone who did the job by hand sees it.
so the scarce skill becomes reviewing, and reviewing needs a real mental model of the job.
for builders: before you give a process to an agent, write down how it's done and who checks it. if you can't, the agent won't fix that for you.
Five Bots now have full-time roles on our accounting team. We onboard, train, and manage them just like human teammates.
Here's how we built them to handle the order-to-cash process, from provisioning customer accounts to collecting payments.
x.ai/news/grok-bot-accountin…
8
717
I was said when I say @julesagent going away.
But getting the loops back in Stitch is interesting.
Made my setup, started to create loops, added context, added specs.
Now waiting for what it will create.
Let's build more!
Introducing Stitch Loop, a new Stitch experiment.
Stitch helps you design your product. Stitch Loop helps you keep improving it once it's live: a product agent built for outcomes, not just PRs.
Stitch Loop monitors your production data and lets you build loops that work continuously until your goals are met.
You can define your own custom loops, or start improving your product on day one with built in Loops for:
🛡️ Security: built with Gemini 4 Carbon and @GoogleDeepMind CodeMender
⌚️ Latency: to keep your product fast
👋Accessibility: so it works for everyone
Join the waitlist: stitch.google.com/loop/waitl…
Learn more 🧵👇
1
10
866
things are moving fast.
private LLMs, owning models and data is getting important.
not only for regulated businesses, but like any startup too.
AI is opening up a lot of new things.
open-weight models closed the quality gap.
the new bottleneck is serving: cost per token, latency under load, and where your data is processed.
Alpha has joined NVIDIA Inception to scale that layer.
alpha.sh/blog/alpha-joins-nv…
a draft ERC on Ethereum asks a good question: when an issuer changes a token, where is the "why"?
today a rescale, a new issuance or a parameter change shows up as raw event logs. no reason, no link to the disclosure behind it.
the proposal: an Announcement event opens a bracket, the state changes run inside it in the same transaction, an EndAnnouncement closes it. one id ties the change to a document, and later corrections keep the same id.
for tokenized real-world assets this matters. a corporate action gets announced, corrected, confirmed, then executed. the chain only sees the last step.
the same gap exists for issuer actions on MultiversX tokens. a multisig / governance contract can be set as the token controller and these implementations can be done there.
builders: would you want this as a standard, or is a link in the transaction data field enough?
7
726
xExchange merged a small PR today that every contract team should copy: reproducible release builds.
the problem: you read the source on GitHub, but what runs on chain is a .wasm file. how do you know one came from the other?
what mx-exchange-sc #1081 does:
- release builds run inside a pinned Docker image, sdk-rust-contract-builder v12.0.0
- the reproducible-build workflow from mx-sc-actions is pinned to one commit
- it runs by hand from the release tag, for one contract or all of them
- the release notes get the image and the codehash of each .wasm appended
their test release for token-unstake: the wasm built in CI was identical to a local build. same codehash.
that line is the point. same source + same pinned toolchain = same bytes. anyone can rebuild it, hash it, and compare with the code hash of the deployed contract.
two details I like:
- the CI action is pinned to a commit, not a branch. a moving CI dependency is a supply-chain hole
- the build starts from the tag, by hand. a release is a decision, not a side effect of clicking publish
if you ship contracts on MultiversX: could you rebuild your deployed wasm byte for byte today?
"RL generalizes, SFT memorizes" became a rule of thumb after DeepSeek-R1. this paper checks whether the comparison was ever fair.
the setup: Qwen2.5-7B and Qwen3-8B-Base, trained on DAPO math problems, then tested on reasoning tasks outside math.
what they changed on the SFT side:
- the source. a mix of answers from the RL-tuned model and from Qwen3-14B beat answers from the bigger model alone. the best mix in their runs was 70/30
- the scale. RL sees many rollouts per prompt, SFT usually gets one answer per prompt. match the number of correct rollouts and SFT generalizes much better. match the total rollout budget and it lands level with RL
- both together: SFT ends up ahead of RL
the why is the part I liked. they had an LLM label the solution methods in the training data. bigger datasets cover more rare methods, and those rare methods are the ones used on the out-of-domain questions.
honest limits: 7-8B models, math as the only training domain, and the theory part is on shallow transformers, not big ones.
my takeaway: before blaming the algorithm, check how much data it saw and how varied it was.
if you fine-tune: one answer per prompt, or many?
Great paper on why SFT can generalize just as well as RL if you give it enough diverse training data.
Like everyone else since DeepSeek-R1 dropped, I’ve been treating "RL generalizes, SFT memorizes" as gospel.
Whenever SFT runs plateaued on out-of-domain reasoning, everyone just blame the algorithm.
Turns out, we might have been running an unfair comparison the entire time.
In standard setups, RL gets to explore a huge rollout budget per prompt while SFT gets trained on a tiny slice of data.
So the authors decided to actually level the playing field on Qwen2.5 and Qwen3 models by testing two things: mixing data sources instead of just distilling from one giant teacher, and scaling up the SFT data to match RL's rollout budget.
The results completely flip the usual narrative.
A carefully mixed dataset easily beats single-teacher distillation, and matching RL's rollout budget lets SFT generalize just as well as RL.
And when you combine both, SFT actually outperforms it.
To figure out why, they used LLM annotations to inspect the rollouts and proved (both empirically and theoretically on shallow transformers) that larger data budgets expose SFT to rare, "tail" solution methods.
Those uncommon reasoning paths are where the actual generalization signal lives.
Honestly, for everyone tired of debugging finicky online RL pipelines, this is a huge relief.
SFT wasn't broken, we were just starving it of diverse rollouts.
Full paper PDF: arxiv.org/pdf/2611.14912
4
590
"who runs this thing?" is a fair question. the answer just isn't the one a politician expects.
on MultiversX nobody signs off on your transaction. validators run the same protocol rules, every block is checked by a group of them, and a block that breaks the rules gets rejected.
changing the rules themselves goes through on-chain governance proposals.
so the honest answer to "who runs it": a lot of independent operators, all running the same rules.
harder to fit on a TV panel than a name.
So Boris Johnson continues to get Bitcoin wrong, so I set the record straight at Solana Breakpoint.
"Who runs this thing? Yakamoto Nakasori… I don't know where to find him."
- No one has to guarantee it. That's the beauty of the protocol
- Finite issuance, secured by hardware and electricity
- Drug dealers? They'd rather use cash in hand than a traceable blockchain
Healthy scepticism is fine, Boris. But we've got to do better at educating politicians and influencers.
12
863



















