Babak Farhang

Babak Farhang

@gnahraf.bsky.social

applied & engineering physics guy (Cornell) turned programmer Building something new. Here's the first part.. https://crums.io Working deck.. https://docs.google.com/presentation/d/1xDiIncACo92nclaHHeyOikr2TSSpizVuXU21CyADu4M/edit?usp=sharing Aurora, CO

177 followers160 following505 posts

This 👇 Surveillance is not just bad; it's unworkable. Every new phone is riddled with apps that add no (minus) value, whose primary purpose is to surveille you. An AI assistant in background with God-unknown privileges. A security nightmare. And my father must learn to navigate this shit in his 90s

These are the technopolitics of *surveillance*, not ads. There is nothing about advertising that demands surveillance. For centuries, advertising was not meaningfully linked to surveillance. The decision to fuse surveillance with advertising was just that - a *decision*, not an inevitability.

@bsky.app historical art on wikipedia is NOT adult content. The labeling engine routinely mislabels wikipedia_dot_org content "adult". Never encountered a case where it can be genuinely argued it could be. Add a rule, whitelist the site pls: many false positves outweigh any potential false negative

Babak Farhang@gnahraf.bsky.social

Art historians, will note: the last frame "5 cats dancing in hell" is a juicy inversion of Matisse's Dance paintings.. en.wikipedia.org/wiki/Dance_(...

Should LLM's be viewed as prompt-decompression machines? The point of this viewpoint would be to develop a yardstick by which to measure a product's "originality", (the human hand) in a possibly agent-involved production. Something like a fuzzy attempt at reverse-engineering the input prompts.

Here's why this may be possible in principle.. An LLM is a compression of its training data, and assuming, for sake of argument, that its output is deterministic, then the (Shannon, not Kolmogorov) entropy of its output is bounded by the entropy of the input prompts. Implementation-wise..

implementation-wise, we likely would not be searching for an exact prompt sequence that faithfully recreates the product. That would be infeasible. Rather, it might be more fuzzy.. how difficult (or easy) is it to regenerate something "close" (given the model).

Babak Farhang reposted

Ok, this "there are two separate brains" thing appears to be making the rounds, and I want to be absolutely clear: this is a press office overhyping some perfectly reasonable science that won't come as a surprise to most developmental biologists.

Superb high level walk-thru of what LLMs do: the surprise they do so well, but also their inability to handle surprises. I'd call the ability to handle surprises, "real-time learning". AI agents (LLMs harnessed to external programs) can't.. bsky.app/profile/gnah... Cory explains it way better 👇

I'll to come right out and say it: statistical extrapolation is *fine*. One of the most useful ways to understand the present and anticipate the future is measuring the things that happened in the past, find the correlations among them, and extrapolate likely future outcomes from those correlations.

LLMs generate Math proofs much the same way that they generate software. To succeed at either task, they rely on external tools for a host of sub-tasks, even basic arithmetic, and ultimately a compiler to check their work. They typically get plenty wrong at each try, but eventually get it right.

To be fair, a human also uses external tools. But what's missing in LLMs is an ability to *learn, their "context window" is bounded. And cognition *requires an ability to learn, to light a light bulb, to adjust the weights, as it were. LLMs can't. Why the superintelligence threat then is fantasy.

LLMs generate Math proofs much the same way that they generate software. To succeed at either task, they rely on external tools for a host of sub-tasks, even basic arithmetic, and ultimately a compiler to check their work. They typically get plenty wrong at each try, but eventually get it right.

Whether this combination of LLM / external-toolset can be called cognition, for me at least, is settled. No. The agent does surprise on occasion, eg w a surprising insight exhibiting deep understanding. But then moments later it'll disappoint, reminding I've anthropomorphized a chatbot's epicycles.

Re AI and the regulatory capture angle.. A lot of people this side of the pond seem not to have gotten the memo: The US has lost international standing. We can regulate the fuck we want.. the world has moved on. To wit, Microsoft Office, EU, Canada, saga of an ICC judge losing access, on and on.

To be clear, Social Security is not a tax; it's a benefit you pay into until you retire. For years, the federal government counted & used social security surpluses to offset deficits. Now that that surplus is gone (for purely demographic reasons), they want you to forget it always was *your money.

Just to be clear, Social Security is not accounting fraud (obviously). Its structure has to be adjusted over time for changing demographics, of course, but it's a completely sustainable approach to income security for the elderly.

Cloud services, generally. Digital sovereignty is now also a serious biz risk. "Canadian Cloud" (AWS style) could do well I think. Geographic proximity suits CCs well for serving US markets. If combined with a sensible legal regime (not subservient to a capricious US) a serious competitive advantage

Under market conditions, the decay of these American platforms would invite competitors from around the world. The fact that Apple and Google extract 30% of every dollar spent in their app stores would bring forth new app stores who were willing to give better deals to app makers and app users. 40/

Babak Farhang reposted

The only meeting Hakeem Jeffries should be having with Jared Kushner is to tell him he’ll be subpoenaed by the House for a corruption investigation on day one of Democratic control of the House. This is just pathetic stuff from the House Dem leader. www.nytimes.com/2026/08/23/u...

nytimes.comJeffries and Kushner Meet Privately as Midterm Attacks FlyPresident Trump’s son-in-law and the man in line to be speaker of a Democratic-led House discussed how Democrats and the administration could work together.

A note to our meek news editors: Stop it with *potential conflict of interest. It's just plain "conflict of interest". It already contains potentiality (potential corruption). Don't be a redundant suckup.

Minefield? I've never heard of criming (bribes, extortion, and shakedowns) called this way. Do they mean a legal minefield? from the criminal trespasser's point of view?

Eric Lipton NYT@ericlipton.nytimes.com

Trump Paid $2 Million by South Korean Company Facing Trade Investigation The payment illustrates the minefield Mr. Trump has created by maintaining personal financial ties with foreign businesses while he is in office. (July 2026) www.nytimes.com/2026/07/14/u...

What a great project! I've had this crazy idea about animating how my cryptographic hash proofs work (eg crums.io/api/state ). I coded my scheme in java.. JMathAnim will fit my existing data structures. Should work great for static explanatory vids. If it had a JavaScript conversion lib, otoh 1/2

crums.iocrums.io
Frank Delporte@frankdelporte.be

New JavaFX In Action 27 is live! 🎉 @davidcalculin.bsky.social built JMathAnim, a JavaFX library and UI for mathematical animations: LaTeX, fractals, geometrics, video export,... Built by a mathematician who "just needed the tool." 👉 webtechie.be/post/javafx-... #JavaFX #Java #OpenSource

If it had a JavaScript conversion lib, otoh.. it could support a *live* data driven animation. Unfortunately, I'm not an expert JS programmer. Still, if I pair-programmed with an coding agent, I might have a shot. 2/2

A government by corporations for corporations. Everywhere, from municipal services, local gov't, to state and federal.. Wherever we contract out basic services to monopolized, for-profit, private companies, citizens lose autonomy and become their bitches. 1/2

They're bankrupting us. They know it. And they've always known this doesn't end well. Why their leaders are moving to Argentina. (Looking at you Mr. Thiel.) 2/2

Keep leveling these thru November. It's perhaps the one issue most voters agree on. It can gather wind.. winds of change be a howling

Scott Dworkin@dworkin.bsky.social

🚨 BREAKING ACTION: Congress just formally leveled six Articles of Impeachment against Supreme Court Chief Justice John Roberts. From rewriting the Constitution to shield the executive branch to massive conflicts of interest, the high court has gone rogue. Tell your reps to back impeachment NOW: 👇

To all would be investors, xAI, $SPCX bag holders.. Agents are quickly becoming commoditized; the only moat that could possibly justify valuations is Liability/Trust. Trust your IP will not be stolen, liability coverage in case you're the one stealing. I.e. Fiction. venturebeat.com/technology/x...

venturebeat.comXiaomi's new open source, agentic AI coding harness MiMo Code beats Claude Code at ultra-long, 200+ step tasksThe persistent memory system addresses a real and widely felt pain point in agentic development workflows — one that competitors are also racing to solve.

More and more interested in agentic coding that _reduces_ code. A good part of software maintenance involves deduplication, reorganizing, throwing away code. I pair-program (w/ claude) to help with such "reduction" tasks. But it feels less like coding.. more like being in a meeting all day, IIBH

A new foundational method for zero-knowledge proofs. I have read (elsewhere) current non-interactive zk proofs are provably flawed: this method seems to shows a path forward. Really well written article 👇

Quanta Magazine@quantamagazine.org

Shafi Goldwasser (left), Silvio Micali (right), and Charles Rackoff devised a way to prove that a statement is true without revealing anything about why. www.quantamagazine.org/how-unknowab...

"CNBC spoke with 19 executives and investors at HumanX, some of whom asked not to be named in order to speak freely." Why on earth would peeps speak freely only if anonymous? Me thinks it's cuz Sam Altman's investors, having gone all in and bet the house on the wrong horse, are now his henchmen.

Techmeme@techmeme.com

Takeaways from HumanX, one of the AI industry's main events: Claude Code dominated the conversation, while some execs noted China's lead in open-weight models (Ashley Capoot/CNBC) Main Link | Techmeme Permalink

@zhitzig.bsky.social re how to deal w/ AI, I think knowing who (or what) created a piece of content (and when) can help. I've been working on a commitment scheme (and easy-to-use software SDK) for journals and ledgers. Can double as an attribution engine. More info linked from my profile.

Great piece about the return of Big Oil to Washington. And an amusing behind-the-scenes thread about investigating the story

Eric Lipton NYT@ericlipton.nytimes.com

Today, in the Sunday print NYT, we published an investigative look at Harold Hamm, the 13th child of Oklahoma sharecroppers who started an oil company named Continental, which turned him into a billionaire, powerbroker and friend of Donald J. Trump. You may never have heard of Hamm. But you should

Totally normal, well adjusted, human being in the foreground, just a little mad (a lot) at the fellow on the floor for interrupting his presentation.

Aaron Rupar@atrupar.com

incredible photo that's definitely worth at least 1,000 words from Andrew Harnik of Getty

Babak Farhang reposted

This morning, the Trump Administration’s Department of War gave me an ultimatum: call up your troops, or we will. It is absolutely outrageous and un-American to demand a Governor send military troops within our own borders and against our will.

If language itself is a live creature that we collectively host (many hypothesize it is), then LLMs are *foreign hosts, viruses, and it shouldn't surprise if it were to deploy antibodies to kill them.

The Age of Reason Strikes Back Meanwhile peoples across the galaxy were beginning to send mathematicians as their representatives. It began with a widely followed representative of the supreme being, then the Romanians ..

Unlike with ordinary assets, the OCC has no real power over #blockchain (nor does any other entity or gov't body). It can regulate all it wants, but the transactions it deems illegal can never be undone. "Regulating" this offers little but expose the US financial system to new forms of credit risk.

Techmeme@techmeme.com

The US OCC says banks can buy and sell customers' crypto assets on their behalf and can outsource crypto custody and trade execution services to third parties (Danny Park/The Block) Main Link | Techmeme Permalink

Babak Farhang reposted

NY Times Editorial Board: “The first 100 days of President Trump’s second term have done more damage to American democracy than anything else since the demise of Reconstruction. Mr. Trump is attempting to create a presidency unconstrained by Congress or the courts…” www.nytimes.com/2025/05/01/o...

nytimes.comOpinion | Fight Like Our Democracy Depends on It (Gift Article)The patriotic response to today’s threat to American democracy is to oppose Trump soberly and strategically.