Michael Kalus posted a photo:

Michael Kalus posted a photo:

Welcome back to the Abstract! Here are the studies this week that took a bite out of life, appealed to the death drive, gave a yellow light to the universe, and produced hitherto unknown levels of cute.

First, it’s the most epic ocean battle: orcas versus sharks (pro tip: you don’t want to be sharks). Then, a scientific approach to apocalyptic ideation; curbing cosmic enthusiasm; and last, the wonderful world of tadpole-less toads.

As always, for more of my work, check out my book First Contact: The Story of Our Obsession with Aliens, or subscribe to my personal newsletter the BeX Files. 

Now, to the feast! 

I guess that’s why they call them killer whales

Higuera-Rivas, Jesús Erick et al. “Novel evidence of interaction between killer whales (Orcinus orca) and juvenile white sharks (Carcharodon carcharias) in the Gulf of California, Mexico.” Frontiers in Marine Science.

Orcas kill young great white sharks by flipping them upside down and tearing their livers out of their bellies, which they then eat family-style, according to a new study that includes new footage of these Promethean interactions in Mexican waters.

“Here we document novel repeated predations by killer whales on juvenile white sharks in the Gulf of California,” said researchers led by Jesús Erick Higuera Rivas of the non-profit Pelagic Protection and Conservation AC. 

“Aerial videos indicate consistency in killer whales’ repeated assaults and strikes on the sharks,” the team added. “Once extirpated from the prey body, the target organ is shared between the members of the pods including calves.”

I’ll give you a beat to let that sink in, like orca teeth on the belly of a shark. While it's well-established that orcas are the only known predator of great white sharks aside from humans, the new study is only the second glimpse of killer whales targeting juvenile sharks. 

This group of orcas, known as Moctezuma’s pod, has developed an effective strategy of working together to flip the sharks over, which interrupts the sharks’ sensory system and puts them into a state called tonic immobility. The authors describe the pod’s work as methodical and well coordinated.

“Our evidence undoubtedly shows consistency in the repeated assaults and strikes, indicating efficient maneuvering ability by the killer whales in attempting to turn the shark upside down, likely to induce tonic immobility and allow uninterrupted access to the organs for consumption, " the team said. Previous reports suggest that “the lack of bite marks or injuries anywhere other than the pectoral fins shows a novel and specialized technique of accessing the liver of the shark with minimal handling of each individual.”  

An orca attacking a juvenile great white shark. Image: Marco Villegas 

Sharks, by the way, do not attack orcas. Just the opposite. As you can imagine based on the horrors you have just read, sharks are so petrified of killer whales that they book it whenever they sense a nearby pod.

“Adult white sharks exhibit a memory and previous knowledge about killer whales, which enables them to activate an avoidance mechanism through behavioral risk effects; a ‘fear’- induced mass exodus from aggregations sites,” the team said. “This response may preclude repeated successful predation on adult white sharks by killer whales.”

In other words, if you’re a shark, one encounter with orcas is enough to make you watch your dorsal side for life—assuming you were lucky enough to escape with it. 

In other news…

Apocalypse now plz

Albrecht, Rudolf et al. “Geopolitical, Socio-Economic and Legal Aspects of the 2024PDC25 Event.” Acta Astronautica.

You may have seen the doomer humor meme to “send the asteroid already,” a plea for sweet cosmic relief that fits our beleaguered times. As it turns out, some scientists engage in this type of apocalyptic wish fulfillment professionally. 

Planetary defense experts often participate in drills involving fictional hazardous asteroids, such as the 2024PDC25, a virtual object “discovered” at the 2025 Planetary Defense Conference. In that simulation, 2024PDC25 had a possible impact date in 2041.

Now a team has used that exercise as a jumping off point to explore what might happen if it hit even earlier, channeling that “send the asteroid already” energy.. The researchers used this time-crunched scenario to speculate about the effect on geopolitics and pivotal events, such as the 2028 US Presidential elections.

“As it is very difficult to extrapolate from 2025 across 16 years in this ‘what-if’ exercise, we decided to bring the scenario forward to 2031 and examine it with today’s global background,” Rudolf Albrecht of the Austrian Space Forum. “Today would be T-6 years and the threat is becoming immediate.”

As the astro-doomers would say: Finally some good news.

Big dark energy

Son, Junhyuk et al. “Strong progenitor age bias in supernova cosmology – II. Alignment with DESI BAO and signs of a non-accelerating universe.” Monthly Notices of the Royal Astronomical Society.

First, we discovered the universe was expanding. Then, we discovered it was expanding at an accelerating rate. Now, a new study suggests that this acceleration might be slowing down. Universe, make up your mind!

But seriously, the possibility that the rate of cosmic expansion is slowing is a big deal, because dark energy—the term for whatever is making the universe expand—was assumed to be a constant for decades. But this consensus has been challenged by observations from the Dark Energy Spectroscopic Instrument (DESI) in Arizona, which became operational in 2021. In its first surveys, DESI’s observations have pointed to an expansion rate that is not fixed, but in flux.

Together with past results, the study “suggests that dark energy may no longer be a cosmological constant” and “our analysis raises the possibility that the present universe is no longer in a state of accelerated expansion,” said researchers led by Junhyuk Son of Yonsei University. “This provides a fundamentally new perspective that challenges the two central pillars of the [cold dark matter] standard cosmological model proposed 27 years ago.”

It will take more research to constrain this mystery, but for now it’s a reminder that the universe loves to surprise.

And the award for most squee goes to…

Thrane, Christian et al. “Museomics and integrative taxonomy reveal three new species of glandular viviparous tree toads (Nectophrynoides) in Tanzania’s Eastern Arc Mountains (Anura: Bufonidae).” Vertebrate Zoology

We’ll end, as all things should, with toadlets. Most frogs and toads reproduce by laying eggs that hatch into tadpoles, but scientists have discovered three new species of toad in Tanzania that give birth to live young—a very rare adaptation for any amphibian, known as ovoviviparity. The scientific term for these youngsters is in fact “toadlet.” Gods be good.

“We describe three new species from the Nectophrynoides viviparus species complex, covering the southern Eastern Arc Mountains populations,” said researchers led by Christian Thrane of the University of Copenhagen. One of the new species included “the observation of toadlets, suggesting that this species is ovoviviparous.”

Note to Nintendo: please make a very tiny Toadlet into a Mario Kart racer.

Thanks for reading! See you next week.

Guest post by Nicholas C. Weaver

The AI companies and related enterprises are currently spending huge amounts in capital expenditures (CAPEX) to build “Gigawatt” class data centers for various AI-related development.

The scale of the investment is so large it is distorting the economy. These massive expenditures, however, are shortly going to prove to be at least a half-trillion-dollar money pit: these massive data centers are simply not fit for purpose.

But to understand why this is a waste of money, it is critical to discuss two technologies: machine learning (ML) in general and large language models (LLMs) in particular.

What is Machine Learning?

Machine learning is a set of general techniques to develop what are, at heart, pattern matching tools: given a sequence of inputs, what is it? Although these techniques are actually very old, it is only in the past decade and a half, with the introduction of modern Graphics Processing Units (GPUs) that ML became feasible for a wide variety of tasks.

ML systems in general work by taking a large amount of “training data”, feeding it into the ML system to “train” it, and, once complete, the ML system performs “inference”: taking a piece of input and saying what type of output it is.

Of course the ML systems do have some drawbacks: they are opaque, they require a lot of computational power, and they can be wrong! So in general, a good recipe for using ML is:

1. You need to build a pattern matcher.
2. You have no clue what to actually look for (because a conventional pattern matcher would be vastly more efficient).
3. When you are done, you still have no clue what to look for (because you could then rewrite it as a more efficient conventional pattern matcher).
4. It is OK to be hilariously wrong a non-zero percentage of the time.

Although seemingly limited, the result can be quite spectacular: the classic ML success story in the past decade and change is speech recognition.

Speech recognition is a hard problem even for humans: how many times have you misunderstood what someone said? It proved even harder for computers until ML systems grew to a practical scale. Now we take tools like Siri for granted.

Siri started out in the datacenter: when you asked Siri for something, the phone would package up your request and send it to one of Apple’s servers, where a powerful computer would parse the speech. Now it largely runs on the phone itself: I can put my phone into airplane mode and have Siri do things because the CPU in my phone includes specialized hardware to make ML run fast.

This works because ML inference may require a lot of math, but the math itself is very regular, effectively multiplying a large number of numbers together. This allows for circuit designs that just place a lot of multipliers and adders together, which by computer standards is highly efficient. Just building lots of math is only part of the story, however, as the math doesn’t have to be very good.

So instead of using high precision math that can represent 18,446,744,073,709,551,616 different numbers (the 64-bit math that most computers normally use), ML inference can use low precision math. Apple’s system largely uses 16-bit math (able to represent 65,536 distinct values). Other ML systems might use 8-bit math (able to represent 256 values) or even just 4-bit (able to represent a total of just 16 values)! Such math means a reduction in accuracy — but if wrong is OK, a little bit more wrong but a lot cheaper is probably OK too.

This means that ML inference is, whenever possible, best done on the edge. Edge computing might have a slightly higher error rate, as models are compressed to be able to fit on a phone, laptop, or desktop computer. In return, however, the costs of the ML system are greatly reduced. Instead of needing an expensive GPU system in the cloud that may cost $10-75 an hour to run, the cost per invocation is effectively zero.

What are Large Language Models?

A large language model (LLM) is a text-based ML system initially trained on effectively all the text the LLM company can download or pirate. These systems require a massive amount of resources to initially train and operate, but do produce a highly convincing obsequious chat bot.

Unfortunately, LLMs in particular have two critical problems: the wrongness inherent in all machine learning systems, and a fundamental inability to distinguish between “data” and “program”.

The wrongness problem is simply how they operate. LLM purveyors talk about ‘hallucinations’, when the LLM produces some wrong output, as just an unfortunate side effect that they are trying to control. This is a lie. Rather, all of a LLM’s output is bullshit in the philosophical sense: statements that are divorced from whether or not they are true or false. The point of a LLM is to output text that looks right given the training data and query, not to produce text that is right.

Even so-called “reasoning” models aren’t. Instead they output a separate set of text that also sounds like reasoning, but isn’t, as witnessed whenever one tries a random variation of the “river crossing” problem.

Compounding the problem of wrongness is a LLM-specific security flaw: they cannot distinguish between ‘code’ and ‘data’ — the instructions to a LLM are simply more text that is fed into the LLM. This means that a LLM will always be fundamentally vulnerable to “prompt injection” attacks, with all protections being potentially brittle hacks that can’t fundamentally eliminate the problem. A LLM can never be used in a context where it can both “receive an input provided by a ‘bad guy’” and where “do something ‘bad’” is unsafe.

This means that LLMs are simply not fit for purpose in many applications. A LLM can’t act as an ‘agent’, reading your email and then doing something in response, because some spammer will instruct your agent to send a purchase to a Nigerian prince’s credit card processing facility.

This makes LLMs a niche technology: very few proposed applications for LLMs are both allowed to be wrong and can operate safely on untrusted input.

Enter the gigawatt datacenter

LLMs and the closely related image generators, unique amongst most ML applications, require staggering resources for both training and inference. These under-construction data centers feature tens of billions of dollars worth of Nvidia processors, drawing billions of watts of power, at an aggregate cost of at least $500B spent to-date by the major players. Yet these data centers are going to prove to be white elephants — because the intended applications simply won’t exist.

Even for these massive models, the process of inference is already best done on the edge. Deepseek showed that very large models can be shrunk in size with remarkably little impact on accuracy, with many other (mostly Chinese) companies following suit. There is an entire community on Reddit, LocalLLaMA, dedicated to running such systems in a home environment. Once again, “a little more wrong at much lower cost” shows its power in the ML space.

So if a company does want to use LLM, it is best done using local servers, such as Mac Studios or Nvidia DGX Sparks: relatively low-cost systems with lots of memory and accelerators optimized for processing ML tasks. It doesn’t take too many users with $200/month subscriptions or too many calls to a $15 per million token API to justify paying for a few $4000 computers instead.

Training may want the large data center, but we’ve long since hit the point of diminishing returns. There is effectively no more text to train on, as even the LLM systems of a few years ago were trained on almost all the coherent text in existence. Similarly, more training, and more training data, will never eliminate the problem of the LLM being occasionally wrong. Instead, the best use of training is probably specialization: taking a large model designed for general input and shrinking it down for a specific task.

The current models were trained, and can be specialized, without the presence of the half-trillion-dollar worth of money pits that major companies like Amazon, Microsoft, Alphabet, and Meta are building.

DeepSeek v3’s claim of $5M for training is probably an exaggeration — but even 10x the cost would be just $50M in computing time for a model that is fully competitive with the best from OpenAI and Anthropic. DeepSeek v3 is a particularly large model, with 600 billion parameters, but can still run on a cluster of just 8 Mac Minis for inference, and could probably be specialized in a couple of hours running on one or two big servers, at a rental cost of a few hundred dollars.

The gigawatt data center is an evolutionary dead end. When the AI hype-bubble bursts, they are going to be multi-billion-dollar white elephants full of chips that are best simply turned off and the money regarded as wasted. There are not going to be customers: the existing resources are already sufficient for the proposed applications that have a hope of actually working where the rubber meets the road.

The future of ML is at the edge.

I did an interview with Nick about his ideas here! You can watch it on video (360p Zoom, baybee) or listen to the podcast! The transcript is on the Patreon for $5-and-up patrons. [YouTube; podcast]