iiinova — where your digital self keeps evolving

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 15/07/2026

This is "good guy with a gun", but #AI right? "The only way to stop AI risk is more AI"

Screenshot of a website that reads "Protect Your Data From AI, With AI
In a world that runs on data, cut AI risk, strengthen cyber posture, and accelerate your business with data-centric security." A button to "Request a Demo" is present.

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 13/02/2026

5/ 🌍 These pLMs kind suck on generalizing to real-world tasks, data leakage or not. Both pLM based and classic models stumble on: - Human–SARS CoV 2 interaction predictions - Point mutation effects on binding affinity

A Receiver-Operator Characteristic plot of methods w/ and w/o data leakage when tested on human v. human interactions and human v. SARS-CoV-2 interactions. The human-virus interactions have AUROCs close to 0.5 whereas the human-human interactions have good performance.

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 13/02/2026

4/ The pLM data leakage problem isn't present for all tasks. When we tested the same pLMs on protein keyword annotation (a non paired task), performance stayed stable. Leakage might be specific to paired sequence problems like PPI.

A clustered bar chart with keyword categories on the x-axis and performance on the y-axis (Normalized Discounted Cumulative Gain). It shows little or no difference between models with and without data leakage.

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 13/02/2026

3/ 🐇 We came up with a new ultra-fast pLM with no data leakage during these experiments. It's truly one of the more exciting aspects of this article in my mind. Training all these language models was tough work! That's why we created SqueezeProt, a new pLM which […] [Original post on cosocial.ca]

A scatter plot showing inference speed on the x-axis and predictive performance on the y-axis. SqueezeProt-U50 is on the far right, performing 5x faster than the competition while maintaining most of the performance.

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 13/02/2026

2/ 🚨 Data leakage turns out to be a major problem for pre-trained pLMs on some tasks. We found that existing pretrained pLMs unintentionally leak information from training data into downstream PPI tasks, inflating test scores. Data leakage can cause performance scores to increase by over 10 %!

Graph showing performance of PPIs w/ and w/o data leakage. The difference is as high as 11.6%.

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 13/02/2026

🧵 1/ 🎉 So happy to share our latest work that I authored with my #PhD supervisor Amin Emad. While #ChatGPT and #Claude are natural language models (#llms), computational biologists have been using protein language models (pLMs) which model protein sequences. We […] [Original post on cosocial.ca]

A diagram showing how PPI testing proteins are present in upstream pre-trained pLMs.

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 09/01/2026

When watching a presentation with a slideshow: do you like textual description of figures/data or would you rather just see the figures/data and hear the presenter describe it. #poll #presentation #slideshow

Slide with a scatter plot on the left and on the right a figure title in bold and a text description below it. The scatter plot shows log parameters on the Y axis, Release date on the X-axis, and various Open and Closed LLMs as points.

Slide with a scatter plot centered with it's bold title just above it.. The scatter plot shows log parameters on the Y axis, Release date on the X-axis, and various Open and Closed LLMs as points.

000 0

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 04/11/2025

When an incumbent government blames vote splitting for their loss, don't fall for it. They had the power to enact electoral reform, but they chose not to. It's a power play they all take, and it's costing us a representative #democracy #montreal #cdnpoli […] [Original post on cosocial.ca]

032 3

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 20/09/2025

On the long list of ways in which I wish my #phd supervisor would allow me to end my #thesis #academia #academicchatter

Screenshot which reads "Regrettably, performing such a thorough analysis as would be required is beyond the scope of this thesis and could demand another doctoral degree's worth of study. And as they say, "Fool me once...""

040 4

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 15/09/2025

The example text Claude Shannon uses in "Prediction and Entropy of Printed English" (1950) is probably one of my favourites and people should use it more often. #phd #academia #academicchatter

Screenshot of the aforementioned paper which contains the example text

THERE IS NO REVERSE ON A MOTORCYCLE A FRIEND OF MINE FOUND THIS OUT RATHER DRAMATICALLY THE OTHER DAY

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 10/09/2025

#Zotero has always been but the new versions are just crazy good.

Screenshot of a Zotero tooltip showing a bibliography entry when hovering over its in-text numeric citation.

Demonstrating highlighting in multiple colours (yellow and green) and adding text comments.

000 0

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 02/09/2025

Over-under that this figure is the result of taking a photo of a CRT monitor. From Sander C, Schneider R. Database of homology-derived protein structures and the structural meaning of sequence alignment. Proteins: Structure, Function, and Bioinformatics. 1991 […] [Original post on cosocial.ca]

Fig. 2. Color stereo view. Calibration of the homology
threshold is based on this 3-D scatter plot of sequence similarity (Y. range 0-100%), structure similarity (Z, range 0-100%) and alignment length (X, range 0-150 residues) for pairwise protein sequence alignments. Each point represents the alignment of two protein fragments, one each from a protein of known 3-0 structure produced by a standard sequence alignment method. Red points are pairs dissimilar in structure (bad pairs), blue points are pairs similar in structure (good pairs), with other colors interpolating intermediate values of structural agreement. The rectangular blue
slice (back) represents good pairs; they occur at almost all values of sequence similarity and all lengths-the thin population in the top half of the blue slice being an artifact of the database (very few protein pairs with 50-90% sequence similarity solved by crystallography). The absence of (yellow and red) points in the top left and front shows that no pairs with sufficiently high sequence similarity have low structure similarity: above the homology threshold, all points are "pushed" into the blue region of good structural agreement. Sequence-identical oligopeptides (5-10 residues long) with dissimilar local structure6 are red points at the front top right. Homologous protein pairs of length about 150 residues are blue points at the back top left. (alt text truncated)

030 3

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 12/08/2025

Eeek! My first time on the front page of #HackerNews (even if it's only with 4 points). Check out the #blog post! Maybe give it an updoot! jszym.com/blog/archiving_googl

Screenshot of the frontpage of HackerNews, at around 11:25PM EST on August 12th. My blog post is #8 with 4 points.

201 2

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 08/08/2025

@johannab Ah, it seems our provincial flower has taken root in Ontario. Apologies.

Many traffic cones and construction signs in Montreal.

000 0

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 18/07/2025

La ville de #montréal demande des commentaires sur un projet de loi visant à lutter contre les fermetures de salles de concert dues à des plaintes pour nuisances sonores, comme nous l'avons vu à de nombreuses reprises. Si les fermetures de salles de concert vous […] [Original post on cosocial.ca]

Je suis profondément préoccupé par la fermeture d'importantes salles de concert à Montréal en raison de plaintes abusives liées au bruit, qui modifient rapidement et pour le pire l'esprit et la culture du centre-ville.

Je demande que des mesures soient prises rapidement. La semaine dernière, nous avons vu la fermeture du Barfly sur St-Laurent en raison de plaintes liées au bruit. Notre inaction nous coûte cher.

Il est important de noter qu'un grand nombre de ces plaintes liées au bruit concernent des rues commerciales dynamiques, ce qui soulève la question de savoir s'il s'agit également d'un échec du zonage des immeubles résidentiels dans des zones non appropriées, et si les établissements situés dans des zones commerciales très fréquentées devraient être soumis aux mêmes exigences en matière de bruit que ceux situés dans des zones qui ne le sont pas.

002 2

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 18/07/2025

The city of #montreal is asking for comments on draft legislation to tackle the shut-down of music venues due to noise complaints that we've seen happen over and over. They're doing this by increasing noise thresholds for venues. If the music venue closures […] [Original post on cosocial.ca]

I am deeply concerned with the closure of important music venues in Montreal due to spurious noise complaints which are rapidly changing the spirit and culture of the city's downtown core for the worse.
I ask that action be taken quickly. Just last week we saw the closure of Barfly on St-Laurent due to noise complaints. Our inaction is costing us dearly.
Importantly, many of these noise complaints are on busy commercial streets, and it begs the question of whether this is also a failure of zoning residential buildings where they are not suitable, and also whether venues on high-traffic commercial zones should be subject to the same noise requirements as those in areas which are not.

111 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 16/07/2025

The myoglobin was apparently from a sperm whale. No, I have no idea why.

Screenshot from the paper which reads "The present article describes the application, at low resolution, of the isomorphous replacement method in three dimensions to type A crystals of sperm whale myoglobin."

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 16/07/2025

Before we had 3D generated models of protein structure, we had... whatever this awesome thing is. Left from Kendrew, J.C., Bodo, G., Dintzis, H. et al. A Three-Dimensional Model of the Myoglobin Molecule Obtained by X-Ray Analysis. Nature 181, 662–666 (1958) […] [Original post on cosocial.ca]

A screenshot of pg. 665 from Kendrew, J.C., Bodo, G., Dintzis, H. et al. A Three-Dimensional Model of the Myoglobin Molecule Obtained by X-Ray Analysis. Nature 181, 662–666 (1958). https://doi.org/10.1038/181662a0

The screenshot includes a figure whose caption reads: "Photographs of a model of the myoglobin molecule. Polypeptide chains are white; the grey disk is the haem group. The three spheres show positions at which heavy atoms were attached to the molecule (black: Hg of p-chloro-mercuri-benzene-sulphonate; dark grey: Hg of mercury diammine; light grey: Au of auri-chloride). The marks on the scale are 1 A. apart"

Modern computer-generated 3D model of myoglobin from RCSB PDB.

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 15/07/2025

I always assumed much of our GHG emissions came from industry and energy generation, but wowza... look at how much the oil and gas industry emit (notably, we're not talking about domestic energy generation). It's like the only share of the pie that's grown […] [Original post on cosocial.ca]

This stacked column graph shows GHG emissions in Canada by sector from 1990 to 2022 in MT of CO2e. Total GHG emissions have increased in Canada from 608 MT of CO2e in 1990 to 708 MT of CO2e in 2022.

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 03/07/2025

Nobody will notice if I leave this in my #thesis, right? #phd

2.1 The Biological Basis of Protein-Protein Interactions

Proteins are the subject, verb, and object of a dizzying number of molecular functions that make-up life. The structural organization, metabolism, proliferation, motility, and death of a cell are just a Costco sample of the bulk-discounted 20 pound frozen-packed pre-cooked tangle of angel hair pasta that we call molecular pathways.

130 3

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 04/06/2025

The best part of running a local #llm is getting to change the default prompt instructions.

Screenshot that reads: "Would you like me to run this code to verify it works as expected, Big Chief?

4.02 tok/sec • 959 tokens • 3.09s to first token • Stop reason: EOS Token Found"

001 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 22/05/2025

Today's edition of "amusing out-of-context figure captions". #science

A graph with DNA (mg) on the y-axis and "Months" on the x-axis. Caption reads. Fig 4. Each point represents a single brain.

000 0

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 29/04/2025

This (deeply flawed) model, is increasingly certain that we're going to have an LPC minority. #canada #cdnpoli #elxn45 #jszym_elxn45

SEATS
------
🔴 LPC 168±2
🔵 CPC 148±2
⚜ BQ 22±1
🟠 NDP 4±0
🟢 GPC 1±0
🟣 PPC 0±0

OUTCOMES
------
lpc maj 3.4%
lpc min 96.6%
cpc maj 0.0%
cpc min 0.0%
other 0.0%

110 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 29/04/2025

With 51% of the vote already counted, Pollievre remains behind his LPC counterpart in his riding of Carleton. #canada #cdnpoli #elxn45 #jszym_elxn45

Carleton
Ontario
Polls reporting: 175 / 266

LIB
Bruce Fanjoy
51%
21,584
CON Incumbent
Pierre Poilievre
45.5%
19,257
NDP
Beth Prokaska
1.5%
650
GRN
Mark Watson
0.6%
270
UPC
Karen Bourdeau
0.1%
49

Leader riding
Conservative leader Pierre Poilievre is up for re-election in this riding, which he has represented since 2004. There are 91 candidates in the riding after a group called the Longest Ballot Committee signed up dozens of Independent candidates as part of a recurring protest against the first-past-the-post voting system, which is expected to prolong counting in the riding.

111 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 29/04/2025

The model has been updated! The model, which remains too bullish about the LPC's fate, is now pointing heavily towards a LPC minority. #canada #elxn45 #jszym_elxn45 #cdnpoli

SEATS
------
🔴 LPC 170±2
🔵 CPC 146±2
⚜ BQ 21±1
🟠 NDP 4±1
🟢 GPC 1±0
🟣 PPC 0±0

OUTCOMES
------
lpc maj 28.8%
lpc min 71.2%
cpc maj 0.0%
cpc min 0.0%
other 0.0%

130 3

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 29/04/2025

Now that we have a few seats called, here is the updated projection. This (bad) model is still saying 100% LPC majority. Note: the model take 30 mins to run so it's on a delay. #canada #elxn45 #jszym_elxn45 #cdnpoli

SEATS
------
🔴 LPC 189±4
🔵 CPC 133±4
⚜ BQ 19±2
🟠 NDP 1±1
🟢 GPC 0±1
🟣 PPC 0±0

110 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 29/04/2025

These are the 20 ridings I'm looking most closely at. Only four are west of Ontario, which, as in most elections, we're going to probably know the fate of the election by the time Ontario is called. The columns here are win probabilities reported by the model […] [Original post on cosocial.ca]

A list of 20 ridings that are the most competitive.

420 4

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 28/04/2025

I made a horribly hackey and quick election forecast model last night for #elxn45 and this is what I came up with. It doesn't take into account demographic or candidate adjustments, and it's almost certainly underestimating NDP and GPC seats, but it backtests […] [Original post on cosocial.ca]

SEATS
------
LPC 191±4
CPC 132±4
BQ 19±2
NDP 1±1
GPC 0±1
PPC 0±1

221 2

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 17/04/2025

This is like the rise of Third Way parties after Thatcher/Reagan, but compressed on the scale of like 100 years. #uspol #auspol

Quote tweet by @utopia-defer.red on Bluesky which reads "American exceptionalism is electing a pariah government so repugnant that it tanks the electoral prospects of aligned right wing goverments across the entire anglosphere.

The quoted tweet by @geoffdeburca.bsky.social that reads Here's the odss over time in Australia. An attached plot of election odds forecasting showing Labour dramatically rising and Conservatives dramatically falling.

123 3

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 27/02/2025

I have to say as someone who no longer uses #gmail , the old interface was peak internet (minus the ads). #ux

A screenshot from gmail c. 2004. An email chain between two people is viewable.

000 0

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 13/02/2025

Someone needs to be fired for writing "handful of fingertip injuries". #canada #cbc

A screenshot of a news headline that reads "Igloo recalls more than 1M coolers after handle hazard causes fingertip amputations
Igloo is recalling more than 1 million of its coolers sold across Canada, the U.S. and Mexico due to a handle hazard that has resulted in a handful of fingertip injuries, including some amputations." The cooler in question is also shown.

100 1

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 09/11/2024

They really are downplaying the fact that this spider _sometimes_ eats birds.

Screenshot from Wikipedia that reads "The Goliath birdeater (Theraphosa blondi) belongs to the tarantula family Theraphosidae. Found in northern South America, it is the largest spider in the world by mass (175 g (6.2 oz)) and body length (up to 13 cm (5.1 in)), and second to the giant huntsman spider by leg span.[1] It is also called the Goliath tarantula or Goliath bird-eating spider;[2] the practice of calling theraphosids "bird-eating" derives from an early 18th-century copper engraving by Maria Sibylla Merian that shows one eating a hummingbird. Despite the spider's name, it rarely preys on birds."

Screenshot from Wikipedia which depicts the "Goliath birdeater". It's a big tarantula.

000 0

Joseph Szymborski :qcca: @jszym.cosocial.ca.ap.brid.gy · 29/10/2024

I really do hate the #transformers nickname, it's so devoid of information. Strong vote to #retvrn to "self-attention networks". arxiv.org/abs/2006.11316

Footnote 1 from the attached paper that reads: Neural networks that use the self-attention modules of Vaswani et al. are sometimes called "Transformers,"
but in the interest of clarity we call them "self-attention networks."

000 0