@firstdogonthemoon

Love "professional quandongs"!

https://www.theguardian.com/commentisfree/picture/2026/aug/03/its-the-debate-of-the-century-two-climate-denialists-battle-it-out-to-see-who-agrees-the-most-denialistly

I've been sorting the Australian #PlumeMoth specimens in the Australian National #Insect #Collection. #ANIC includes a few specimens collected in the wider region. One of them is this extraordinary plume moth collected by W.W. Brandt in July/August 1958 in Kiambaui / Kiambowe in north-eastern #PapuaNewGuinea. Sadly, it's not properly set, so it's hard to photograph well.

The location is here:

https://www.openstreetmap.org/search?query=kiambaui&zoom=9&minlon=139.39175736604636&minlat=-7.101001739522618&maxlon=146.6839326590151&maxlat=-3.0007513241170676#map=15/-5.70511/146.26395

The basic colour pattern is not uncommon here in #Australia in #Gelechioidea and #Tineidae, but I've never seen any plume moth with markings at all similar.

It will be a new species in the tribe #Platyptiliini, probably somewhere close to #Xyroptila.

#Lepidoptera #Pterophoridae #entomology #moth

5/5

In practice, this role will require at least the following:

1. Protect datasets produced by direct sensing or created by human input.
2. Ensure the provenance of all datasets is clear and that all transformations and modeled steps are thoroughly documented – in general maximise the FAIRness of all data.
3. Avoid all references to undifferentiated “AI” – in all cases, be explicit about exactly what categories of machine learning are used at each stage in any workflow.
4. Carry out a risk assessment for any use of machine learning and document all known biases and uncertainties, including externalities such as uncredited training materials and environmental costs.
5. In any cases where generative ML techniques are used, ensure that the architecture of the tool, the sequence of transformation steps and any intermediate prompts are exposed, documented and, where possible, explained using a deterministic mapping.
6. Ensure that training data for all models is ethically sourced and fully attributed, and where applicable follows the CARE Principles for Indigenous Data Governance.

4/5

The scientific community (including academic and science-based organisations, data repositories and knowledge infrastructures) faces the challenge of how to respond to these trends. (Non-generative) machine learning remains a formidable tool that can advance their interested in well-understood and statistically quantifiable ways. In some contexts, LLMs may also prove valuable, for example in deriving candidate metadata for a dataset based only on text documents and/or the data to be described. However, unless the provenance of these metadata is fully documented along with the fact that they await human verification, even such constrained uses risk polluting the evidence-base for science and for downstream applied uses in policy, society or industry. There may also be beneficial uses for LLMs in translating natural-language queries into structured forms and translating specialised information results back into natural language. In such cases, the risk of non-determinative results could be mitigated by accompanying ouputs with unambiguous presentation of the form of intermediate queries (perhaps with a non-LLM restatement of these queries in natural language).

In a world increasingly filled with information-shaped synthetic outputs, access to trusted and well-documented evidence sources will become even more significant. In some countries, previously accessible public databases are being taken down. Foundational knowledge components across the Internet, including Wikipedia and the Internet Archive, are under threat from political opponents, funding challenges and denial of service attacks. Many trusted sources are paywalling their content or removing it entirely to limit theft and uncredited use. The scientific community urgently needs to recognise that a central and critical part of their future role will be to deliver unpolluted information in forms that all users can trust.

3/5

It must be noted that by default the outputs from generative models for any specific application can only be evaluated on subjective criteria of taste or feel. Understanding the actual trustworthiness of the outputs for a given purpose will in each case require significant investment in defining measurable criteria and establishing benchmarks to repeat testing (since the behaviour of one generation of a model may be different from another).

Across the Internet, there has been a drive led by Silicon Valley to insert LLM-based interfaces into all online platforms. Users are increasingly encouraged to chat with an AI agent rather than being offered direct access to structured knowledge sources. This trend includes the prioritisation within major search engines of “AI answers” over links to sources. The result is that the bulk of information access is mediated through opaque services that have undisclosed algorithms and unmeasurable or unknown success metrics. The ease with which text and imagery can now be synthesised means that the sources feeding the models for these applications are increasingly likely to be compromised with false information, marketing hyperbole and propaganda. Given the opacity of the agentic tools delivering answers to users, the scale and biases of such misinformation cannot be known. These information channels are in the hands of a small number of ultra-wealthy commercially-centred entities. There is accordingly a clear opportunity for these few completely to control public opinion and discourse in all regions to the detriment of democracy everywhere. Even if these corporations were completely benign, the best case is that agentic systems will tend to reinforce biases found in their training data.

In addition to the issues outlines above, generative ML relies on access to massive volumes of (predominantly unrewarded and unacknowledged) digital source materials and an unprecedented expansion in the carbon footprint, water consumption and use of rare minerals by the tech industry. The purveyors of generative ML have articulated their ambition to siphon value from labour and economic activity across all sectors. Regardless of any possible beneficial uses, this technology is a ruthless effort to enrich a few mega-corporations at the expense of all.

2/5

The recent explosion in interest in AI solutions leverages the positive cultural associations from such successful applications but centres specific on generative ML models. These offer streams of text, imagery, audio or video as likely examples of the kind of text, image, etc. that would in some way extend a given prompt (which may again be text, a possibly incomplete image, etc.). In the case of text, the output from a Large Language Model (LLM) is a potentially infinite stream of text tokens that are likely continuations of the prior tokens, based on probabilities derived from massive volumes of existing text. In the case of imagery, new sections of an image are generated again fundamentally through applying learning of what is likely in a given context (either a source image or a text prompt). These tools echo the vocabulary, syntax and associations found in their input data. With suitable training materials, generative ML can not only synthesise new text but predict likely mappings between text in different forms, including between spoken languages, between text and structured data formats, and between text and software code. In each case, at least parts of the process involve generative ML stages to synthesise a form appropriate for the next stage in a chain.

Generative ML solutions underpin most of the current growth in AI applications and data centres. Although the corporations delivering these tools do not share details of all the internals of their models, it is clear that the underlying LLMs and other generative tools are embedded in a matrix of other software that aims to increase the acceptability of the responses. Strategies may include injection of additional prompt information; adjustments to the weights associated with particular sequences of tokens; detection of prompts that can be converted into API calls to more deterministic tools such as calculators or databases; iterative or parallel execution of models to discard outliers; or chaining multiple stages of generative ML processing to form a workflow. As these applications become more complex, they are increasingly marketed as chatbots or as agents and described in terms that suggest autonomous intelligence. In reality, every agentic tool is a proprietary aggregate that combines generative ML with other components (which may be well reasoned and designed or may be ad hoc tweaks) to create a synthetic model that is intended to meet user expectations and needs as often as possible.

I have a couple of contexts in which I need to provide formal input on the use of #AI for #science. I've written the following and would welcome suggestions if I have misstated or misrepresented any aspects of the #genML craze.

1/5

Machine learning (ML) and artificial intelligence (AI) are variably defined overlapping terms that are generally applied to software applications that deliver information outputs that go beyond data captured in prior databases or knowledge systems. The term AI is currently normally applied to applications that simulate the behaviour and responses of human agents.

ML approaches are a significant tool of great importance both for science and applied use. Successful applications have exploited diverse technologies including semantic datasets, process models and, above all, neural networks to estimate unknown values for variables of interest. Such applications include, among others, weather forecasting, future climate prediction, text recognition, audio processing, and image classification (including processing of biomedical imagery or species identification tools). Common features of such applications include well-curated source datasets that include known values for the variable of interest associated with relevant co-variables; prediction or estimation of output values (temperature, species name, etc.) that include significantly less information than the inputs; and well-defined truth criteria that allow the success rate of the application to be quantified. The information produced by such models, when combined with success measurements, is in effect a probability for a given value in a given context based on co-variables and on the training data.

Other successful AI/ML applications such as protein-folding models and computer chess may have very different internal models that have been developed to targeted a specific domain but still have well-defined and measurable success criteria (proximity of modeled protein structures to actual molecules, victory in chess matches).

Tawny #Frogmouth in our front garden in #Canberra this evening.

Checking the #iNaturalist records for Australian #PlumeMoths in the genus #Hexadactilia, it seems there is an outlier that doesn't fit either of the two known species:

https://inaturalist.ala.org.au/observations/285825134

This is further north than the other known records of the genus from #Australia and has rather different markings. It may well be an #undescribed species. The only other possibility is a third species described from #Sabah in #Borneo, but I can't locate a copy of the relevant paper. It's from a 1995 issue of Entomologische Zeitschrift Frankfurt am Main:

Arenberger, E. (1995) Eine neue Art der Gattung Hexadactilia Fletcher, 1910 von Sabah. Ent. Z., Frankf. a. M. 105: 138-140

Does anyone here have access to that volume?

#literature #LiteratureRequest #entomology #Lepidoptera #Pterophoridae #Queensland #taxonomy #biodiversity

The #FieldNaturalists Association of #Canberra has uploaded a video to YouTube of my May talk, Exploring Local Biodiversity with a Malaise Trap:

https://www.youtube.com/watch?v=Z18kSjIslxE

It includes lots of interesting #insects found in my garden and thoughts on undescribed #biodiversity.

#entomology #Australia #wasps #Hymenoptera #moths #Lepidoptera

I've been going through unsorted accessions in the #Australian National #Insect #Collection (#ANIC) here in Canberra to find any plume #moths (#Pterophoridae) and many-plume moths (#Alucitidae) before doing more work to revise the fauna of these two families.

Today, I found this specimen collected by David Rentz in #Kuranda #Queensland in 2007 - it's not a known named Australian species but it looks a close match for Platyptilia microbscura recorded only from #Misool, an island to the west of #NewGuinea:

https://www.papua-insects.nl/insect%20orders/Lepidoptera/Pterophoridae/Platyptilia/Platyptilia%20microbscura.htm

#entomology #Lepidoptera

I've been going through around 5000 #iNaturalist plume #moth records in #Australia that are not yet research grade and for which I'd not yet added identifications. Almost all the species here can be identified well from a good photo.

The most interesting discovery today was a handful of #caterpillar observations that clearly represent the first photos of the #larva of the common species Stangeia xerodes. See the explanation here:

https://inaturalist.ala.org.au/observations/190890238

#entomology #Lepidoptera #Pterophoridae

Today I found six observations in #iNaturalist that include what may well be the only photos ever taken of live many-plume #moths in the genus #Triscaedecia.

This is a rarely recorded genus for which I accidentally became co-author on a 2019 global revision: https://www.researchgate.net/publication/332013222

There are six named species, four first described in the 2019 paper.

The new records include the second known individual of Triscaedecia svetlanae, extending its range from mainland #Malaysia to #Sabah in #Borneo, and the first (five) records of the genus from #Australia (far north #Queensland). The latter are probably Triscaedecia septemdactyla, a species previously recorded from New Guinea and the Bismarck Islands. I need to check the national #insect #collection here in #Canberra to see if there are any existing specimens in the unsorted material.

https://www.inaturalist.org/observations?taxon_id=781080

#Lepidoptera #Alucitidae #entomology #biodiversity

I've got nothing positive to say about the colonialist programme with which he's most associated but #Kipling's short stories are a marvel. His prose is so controlled and modern, he throws us into minimally specified contexts, and he can sketch a character or mood with just a few words.

One example I've always loved is his Dayspring Mishandled (from Limits and Renewals). There's so much going on with this story.

https://gutenberg.net.au/ebooks06/0603781h.html#01

Spending the day proofreading a book, so I'm playing my way through CAN's discography in chronological order: Delay 1968, Monster Movie, Soundtracks, Tago Mago, Ege Bamyasi, Future Days, Soon Over Babaluma, Landed. Probably then The Lost Tapes.

I find it excellent background music. Plenty of variation, lots of repetitive rhythms, incomprehensible lyrics ...

https://en.wikipedia.org/wiki/Can_(band)

The devastatingly beautiful closing lines of O Superman (For Massenet) seem even more terrifyingly devastating given the current incarnation of the military-industrial-cryptobro complex.

So hold me, Mom
In your long arms
Your petrochemical arms
Your military arms
In your electronic arms
Ah, ah, ah, ah

#LaurieAnderson

Now playing:

https://en.wikipedia.org/wiki/Big_Science_(Laurie_Anderson_album)

"This is your captain ... We're going down ..."

I do love living near #Canberra bushland.

Can someone help me understand the choose-one-of-these-two-digit-numbers #Google #2FA process.

I'm trying to help someone complete access to a couple of accounts on a new Android phone.

After entering the password for one of them, it's clear that it is associated with the old dead phone and asks for the correct number to be selected on that phone.

For the other, it identifies the new phone as the one that should offer the numbers to choose between.

But no messages ever appear on the phone. It's connected to the network, but nothing.

I don't really know how these codes are sent to the phone.

Are they dependent on Google Authenticator having been set up, or are they completely separate?

What needs to be correctly set up on the phone to have it show these numbers?

Clearly, this involves some Google account handshaking. Having control of the phone number is not enough. But I can't understand why they are not appearing on the phone.

And yes - it would have been good if recovery codes had been saved, but that's way too tech-savvy for most people.

Slides from my talk last night to the Field #Naturalists Association of #Canberra:

"Lessons from #Malaise trapping in Aranda"

https://codeberg.org/dhobern/presentation/src/branch/main/Lessons_from_Malaise_trapping_in_Aranda.pdf

As with most of my talks, there are few notes, but the gallery of amazing #insects may mean something to some of you.

#entomology #taxonomy #DNAbarcoding #biodiversity #OnceAndFutureKing #Hymenoptera #Lepidoptera #Strepsiptera #Hemiptera #Collembola #Araneae #wasps #moths #springtails #spiders