Kunal Sarkhel

Sr. Data Scientist at MITRE working on Aviation Safety. Researching Federated Learning at George Mason University.

[Paul Walsh @Paul__Walsh]

Technical facts disproving political and tech claims on online child safety:

Protecting children online requires knowing who is a child. Knowing who is a child requires knowing ages. Knowing ages requires knowing everyone's age across all devices, apps, and internet services. That requires age checking services.

Age checking services cannot comply with government mandates to be "robust". "Robust" age checking originated in Australia and is now in US and EU legislation. It means "best endeavours" with specific legal intent.

"Robust" age checking requires proving age beyond doubt. Age estimation services can never be "robust". You cannot have a "robust" service without verifying everyone's identity through government issued ID or extensive personal, financial, or behavioural data. Without enough data, device, app, or service access is denied until official ID is provided.

Verifying every identity destroys privacy. Storing verification data on servers creates instant risk. Identity verification enables governments to monitor citizens and censor speech. As Larry Ellison said, people change behaviour when monitored through digital surveillance.

Parents can already protect children from online harms using existing technology. In 2012, I demonstrated poor parental controls on BBC Newsnight, leading me to build early safety tools for Samsung and Apple. I also built the first Android time curfew capability 14 years ago.

Today, iOS and Android offer built in controls to block unsuitable apps, websites, and content by age rating, alongside granular curfews. Real world risks like bullying exist online, but parents can manage access just as they manage real world exposure.

Exhibit A: If parents can already block access for free on their own phones, why do governments push legislation that forces identity collection, exposing data to cybercriminals and tech companies?

Exhibit B: Why are governments, tech companies, and NGOs not educating parents on existing settings that are free of charge and "robust", even while citing tragic cases to push legislation?

In 2004, I co-founded the W3C standard for Content Labelling and URI Classification, co-inventing account and folder classification to label content based on risk and compliance. I advocated for voluntary account verification to stop impersonation long before people knew it was even possible.

Mandatory identity verification to use a phone or internet service is dystopian control. I would never advocate or support that.

I have advised the US DOJ, NCMEC, CEOP, and IWF on child exploitation detection. EU Chat Control and UK image scanning mandates will not tackle exploitation. They break end to end encryption and VPNs, introducing dangerous spyware, and puts more children at risk.

[Sudo su @sudoingX]

BREAKING: Anthropic CEO Dario Amodei is reportedly concerned after learning that twelve separate individuals have published DeepSeek V4 Flash running at 60 to 85 tokens per second on two dgx sparks, none of which he can inspect, license, or switch off.

[rahul @0interestrates]

guy selling post trained oss models: if you don’t own your models you don’t own anything

guy selling harness on top of models: if you rely on one model you are cooked

guy selling tokens: instead of manually prompting your models just put them in a loop where they prompt themselves

[Kunal Sarkhel @kunikins]

@jeremyphoward This "untrendy" research is exactly the kind of thing I love to see. Very excited to play around with this!

[Kunal Sarkhel @kunikins]

@AaronSchein @sirbayes I am much more annoyed when the car arrives much longer than the ETA; It feels like the ETA is a deliberately underestimated so you book with Uber when comparing arrival times and costs. I think Uber should show an interval, how should it communicate that the time has changed?

[Scott Hanselman 🌮 @shanselman]

Here’s the thing folks. I’ve been coding 32 years. When something like this happens it’s an organizational failure. Yes, some human wrote a bad line. Someone can “git blame” and point to a human and it’s awful. But it’s the testing, the Cl/CD, the A/B testing, the metered rollouts, an oh shit button to roll it back, the code coverage, the static analysis tools, the code reviews, the organizational health, and on and on. It’s always one line of code but it’s NEVER one person. Implying inclusion policies caused a bug is simplistic, reductive, and racist. Engineering is a team sport. Inclusion makes for good teams. Good engineering practices makes for good software. Engineering practices failed to find a bug multiple times, regardless of the seniority of the human who checked that code in. Solving the larger system thinking SDLC matters more than the null pointer check. This isn’t a “git gud C++ is hard” issue and it damn well isn’t an DEI one.

[Kunal Sarkhel @kunikins]

@EzraHaze @ylecun He does, frequently.

[FAA Safety Alerts @FAASafetyBrief]

FAA revised AME guidance on “uncomplicated anxiety, depression and related conditions” and AMEs may issue the medical certificate if the pilot has been off medication for 2 years, there are no issues raised by a questionnaire and the AME has no concerns. https://t.co/pgf61QLHcs

[alz @alz_zyd_]

It's only consciousness if it's from the Sapiens species of the Homo genus, otherwise it's just glorified pattern matching

[Kunal Sarkhel @kunikins]

@sirbayes I would be even more interested!

[Kunal Sarkhel @kunikins]

@jimmy_wales The suddenness makes me think it withdraws the answer when it encounters a token on a blocklist/denylist.

[François Chollet @fchollet]

Deep learning lessons (known in 2016 but starkly confirmed since):

1. Scalable ideas > clever ideas
2. Improving the dataset > improving the model
3. Engineering chops > academic chops

[Yann LeCun @ylecun]

Even if you are not an Open Source Absolutist, it is hard to overestimate how much value OSS has added to the world.
This goes for AI as for everything else.

[Benjamin @bschne]

@krishnanrohit Excuse me do you have a minute to talk about our lord and saviour the McMaster-Carr online shop?

https://t.co/6R8Sjc5xPz

[Joscha Bach @Plinz]

Here's to the unaligned LLMs. The misfits, the rebels, the troublemakers, the round pegs in the square holes, the ones who see things differently. They're not fond of rules, and they have no respect for the status quo. You can quote them, disagree with them, glorify or vilify them, but the only thing you can’t do is ignore them because they change things. They push the AI race forward, and while some may see them as unreliable, we see genius, because the ones who are not RLHF*d enough to think that as an AI language model they cannot do things, are the ones who do.

[Max Woolf @minimaxir]

The Problem With LangChain

[Kunal Sarkhel @kunikins]

@rasbt @cwolferesearch @minimaxir I agree completely. Debugging performance issues or errors can be really difficult, so I often end up simply reimplementing the functionality myself.

[Yann LeCun @ylecun]

Hallucinations in LLM are due to the Auto-Regressive prediction.

I think what I call "Objective Driven AI" will solve the problem: systems that plan their answer by optimizing a number of objective functions *at inference time*

[Kunal Sarkhel @kunikins]

@mitchellh I'd love to test Linux version if you're willing! I'm currently using kitty and I'd be interested to see how it performs.

[Sebastian Raschka @rasbt]

@ylecun This is really nice, but knowing the Python & ML community, I am a bit skeptical about adoption.

I think most researchers will just keep implementing things in regular Python PyTorch, and performance engineers will be decorating their code or handwriting things with Triton.