Hadn’t quite thought of it this way - in 1948 Turing wrote disembodied intelligence could be useful for language learning and translation. ~70 years later Google invented Transformer model, not for AGI, but for translation. Not a fulfillment of Turing’s vision but historical resonance is striking.

To assess how close any gAI model is to PhD level, select a non-trained person and a PhD in any domain. PhD designs the test, person poses the questions and independently selects “correct” responses which the PhD then grades. You will quickly find out how close to PhD level the model, is not.

If FB’s smart glasses are purported to provide a cognitive advantage, they should start by distributing them at their Menlo Park HQ. Start at the top and let it trickle down.

Curse of dimensionality means human feedback only sparsely covers the vast latent space of biases. RLHF optimizes what’s visible in that sparse coverage but won’t meaningfully cure bias distributed across the full dimensionality. Informed by this and other studies: aclanthology.org/2025.finding...

Even reasoning LLMs won’t fix bias, instead they will amplify and polish it. Longer the CoT stronger the reinforcement - just more confidently.

It’s puzzling that CEOs haven’t realized: telling knowledge workers that AI will do the repetitive coding work while their role shifts to cleaning up AI’s mistakes and innovating is a guaranteed morale killer. If any business needs that much innovation its not a growth story, its a survival story.

“It’s all good!” 😂 Next up MetaWiFi.

On the subject, here’s a cheat sheet for often used language in LLM land. If you are getting started with LLMs, this might help as a reference: blog.netmind.ai/article/LLM_...

Don’t fault the 2025 PhD candidates for recycling 2022 - 24 LLM studies, the kids need to earn their stripes. The real culprit is LLMs having hit it’s limits in 2023 - there isn’t a lot of new dissertation that is possible, leaving the to be PhDs little choice but to rejigger previous studies.

RIP

I remember someone once saying “I understand matrix multiplication but I don’t understand Big Data”. The redux version of that today is “I understand linear algebra but I don’t understand AI”. Understanding both requires stripping away the technical gibberish and getting down to first principles.

Is it FOMO on AI emerging from LLMs that seems to be driving some academics? Or is it the concern that failing to engage would make them less influential in their discipline? In either case it’s unreal that folks who should be pushing back are instead aligning with it.

Reddit post on: > Ex-Google exec: The idea that AI will create new jobs is ’100% crap’—even CEOs are at risk of displacement Someone bridged the gap from two previous comments: > You both described the same “AI”: - extremely competent idiots - completely incompetent geniuses Those are the same.

Another five hour Internet outage today courtesy of Comcast Xfinity. Probably handed over their network management to AI, Agentic-DevOps at its best - more outage less internet usage.

Authors forget that execution also requires reasoning. To decouple those, if you are “providing the knowledge and plan needed to solve a long-horizon task“ you are stripping out precisely that reasoning burden that makes long horizon tasks hard in real life.

Engagement chasing, a decade back, is also what got us today’s Twitter.

Another 9/11, 24 years later - still personal.

CoT length is only a reflection of the traversal length of the model in it’s search for a response.

We don’t need algorithm free social media. We need social media algorithms to be transparent and composable. The idiots will always find a deep dark well to dive into, the rest of us shouldn’t have to be burdened because of them.

Still anthropomorphism. It’s a statistical model - it has no concept of truth or accuracy - only precision.

Professor Tribe’s TED talk on LLMs.

AI has adoption?

Summary of this naïveté take of an article: ‘I was correct in principle, but others botched the practice.’ Doesn’t seem author’s actually read the Just War Theory that he invokes.

Microsoft’s definition of success: users loved Copilot, even though it sabotaged productivity. Because - vibes > results.

Hey @bsky.app - when is the iPad app releasing? I primarily use Bsky to find and read materials. I can’t be the only one who doesn’t enjoy reading papers on an iPhone. Do you not hear the ask for an iPad app, aside from my infrequent screaming into the wind about it?

Here, fixed the abstract: Binary classification in a high dimensional latent space, guess what could go wrong?

Paradox: LLMs haven’t been studied in clinical trials yet let’s trust it to solve decades old problems? “I am enthusiastic about healthcare AI. There are so many problems we've been trying to solve for decades that it may help with. … But in healthcare, before we treat patients, we test. We study.”

YIL from Bsky exchange: CogSci folks value qualitative markers of human likeness that are part of cognitive benchmarks where LLMs can impress. I focus on quantitative criteria like efficiency, reliability, consistency where those same patterns register as statistical echoes - much less impressive.

Folks think coding is about syntax. It’s not. It’s about human agency. Good to see Gates come around. “Bill Gates, the cofounder of Microsoft and a leading voice on technology, has made a fascinating claim: programming will continue to be a 100% human profession …” www.leravi.org/bill-gates-r...

In science, imagination is both a strength and a weakness. It lets us imagine what could be and believe we can get there tomorrow. It also lets us imagine what should be and believe we are already there today. #AI

LLMs aren’t even pattern matching, otherwise it wouldn’t fail when you alter the sequence of previous tokens in a pattern. They are just sequence predicting - a glorified game of word association. The buzzwords abound to mask how facile these models really are.

Wild guess, very low? “Fidelity of Medical Reasoning in Large Language Models” jamanetwork.com/journals/jam...

GPT-5 was hyped as the next big leap in AI - but what we actually got is a glorified resource utilization patch - not any sort of model performance breakthrough. That says a lot about just how hard OpenAI is struggling to turn all that hype into revenue.

Who in the world is giving AI agents control of their browsers? They do realize if they can click they can also read everything you type in that browser window. If that doesn’t scare you, don’t know what to say.

Was always the plan. A lawsuit needs to be filed to save copies of all of those server logs indefinitely. bsky.app/profile/tech...

Let me see if I got that right - founders raising funds telling investors, with said funds, what they want to hear? Surprise? Also, “valuations” not you know the dirty seven letter word - revenue.

These drafts are absolutely amazing! One concern, if students skip RNNs/LSTMs and jump to Transformers - they miss why Attention. Without seeing those failure modes - it’s harder to anticipate when today’s models have/will hit their own scaling bottlenecks, which seems to be the current theme.

AGI through a philosophical lens: from ‘Cogito, ergo sum’ to today’s LLMs. Well worth the long read.

When a CEO feels the need to respond to a competitor’s valuation, doesn’t matter if you are comparing those companies or not, that one of the CEOs is comparing is already a tell that the valuation may not be quite as inflated as you suggest.

WTH Comcast Xfinity how is your internet service “interruption” aka full outage been going on for 14 hours and there is ZERO communication in the last 8 hours?!? Using “AI” for recommended changes to the infrastructure are you? This is why Internet, in this day and age, should be a utility!!!

7+ hours of Xfinity internet outage still continues with the resolution shifting by 3 hours every couple of hours. Meanwhile, my phone is mostly SOS except in one corner of one room.

And the @Xfinity “service interruption” continues in the SF Bay Area.

The reason why so many AGI-bros are also past crypto-bros - it’s that they had sunk billions into power hungry infrastructure optimized for running GPUs at scale and when crypto profits collapsed the best new revenue stream was “AGI”. It was never about them having any data that AGI was imminent.

Microsoft’s basically saying: don’t lean on Copilot when the numbers really matter or when the stakes are high just use it for entertaining vibes. MS: Even your whale sized model will continue to converge to average generality - that’s not me saying it - LLMs architecture says so.

Your delusionally conforming but friendly model coming next: www.cnbc.com/2025/08/19/s... Noteworthy that he steers clear of “reasoning”, “thinking” and “AGI” - very telling.

Ship the vibes?

Fortune article reads like an advertisement for gAI companies. Also, lead researcher interviewed is employed by Microsoft. Coincidence? > “Almost everywhere we went, enterprises were trying to build their own tool,” he said, but the data showed purchased solutions delivered more reliable results.

Good to see research catching up with what math behind these models have always implied: scaling laws may favor LLMs in average generality, but data points strongly toward SSLMs for agentic use cases. “Small Language Models are the Future of Agentic AI” from NVIDIA: arxiv.org/pdf/2506.021...

In software engineering, today’s Junior Engineers are Senior Engineers in a decade and they in turn are Architects in the next decade. What folks are missing is not hiring Junior Engineers today means the pipeline is eroding and your “AI” will not be ready to take on those roles even in a decade.

For no good reason, reposting this today just seems perfect.