Reflections from a week of extremely human conversations about AI:

There are a lot of reasons why I believe AI will not replace humans but one of the main ones is that it doesn’t matter how much you have in common with any one person; human interaction almost always leads to non-linear thinking. And that’s largely because while we all have humanity in common we all have a unique set of experiences.

To put it in more machine learning centric geeky terms:
Humans have all the same underlying “statistical model” but we all have uniquely different “training data” and that is what leads to a world where each of our own’s training data leads to us uniquely different “optimization functions”.

Quite frankly that’s what makes startups unique, cities dynamic and things fun.

Looking forward to exploring even more training data in upcoming conversations.

Happy Pi Day! For those who celebrate 🤓!

3.14159265358979323846264338327950288419716939937510582097494459230781640628620899862803482534211706798214808651328230664709384460955058223172535940812848111745028410270193852110555964462294895493038196442881097566593344612847564823378678316527120190914564856692346034861045432664821339360726024914127372458700660631558817488152092096282925409171536436789259036001133053054882046652138414695194151160943305727036575959195309218611738193261179310511854807446237996274956735188575272489122793818301194912983367336244065664308602139494639522473719070217986094370277053921717629317675238467481846766940513200056812714526356082778577134275778960917363717872146844090122495343014654958537105079227968925892354201995611212902196086403441815981362977477130996051870721134999999837297804995105973173281609631859502445945534690830264252230825334468503526193118817101000313783875288658753320838142061717766914730359825349042875546873115956286388235378759375195778185778053217122680661300192787661119590921642019893809525720106548586327886593615338182796823030195203530185296899577362259941389124972177528347913151557485724245415069595082953311686172785588907509838175463746493931925506040092770167113900984882401285836160356370766010471018194295559619894676783744944825537977472684710404753464620804668425906949129331367702898915210475216205696602405803815019351125338243003558764024749647326391419927260426992279678235478163600934172164121992458631503028618297455570674983850549458858692699569092721079750930295532116534498720275596023648066549911988183479775356636980742654252786255181841757467289097777279380008164706001614524919217321721477235014144197356854816136115735255213347574184946843852332390739414333454776241686251898356948556209921922218427255025425688767179049460165346680498862723279178608578438382796797668145410095388378636095068006422512520511739298489608412848862694560424196528502221066118630674427862203919494504712371378696095636437191728746776465757396241389086583264599581339047802759009946576407895126946839835259570982582262052248940772671947826848260147699090264013639443745530506820349625245174939965143142980919065925093722169646151570985838741059788595977297549893016175392846813826868386894277415599185592524595395943104997252468084598727364469584865383673622262609912460805124388439045124413654976278079771569143599770012961608944169486855584840635342207222582848864815845602850601684273945226746767889525213852254995466672782398645659611635488623057745649803559363456817432411251507606947945109659609402522887971089314566913686722874894056010150330861792868092087476091782493858900971490967598526136554978189312978482168299894872265880485756401427047755513237964145152374623436454285844479526586782105114135473573952311342716610213596953623144295248493718711014576540359027993440374200731057853906219838744780847848968332144571386875194350643021845319104848100537061

So the Honey browser extension was basically Adware / Malware that also performed e-commerce fraud? Yikes.

https://youtu.be/vc4yL3YTwWk

The current and future state of AI workloads or at least based on the most recent discussions appear to converge around scaling challenges.

There is an ongoing debate about whether pre-training of large language models is facing a deceleration in improvement rates, possibly due to limitations in available data and optimization techniques. However, post-training and run-time inferencing continue to show progress, driven by improvements in data quality and memory efficiency.

As AI infrastructure evolves, there is a shift towards smaller, more efficient models, which can operate effectively on less powerful hardware. This is a good thing. While some companies are focusing on large cluster sizes for pre-training, the majority are trying to explore alternative architectures for inference to handle expanding context windows. This would allow for “less prompting” and even greater accesibility by all users (both consumer segment and enterprise segment).

My two cents are that the average enterprise will have to focus more on investing in better, more context aware agents and care more about inference than training expanding the demand for new hardware alternatives. We could all benefit from more optimization and openness

#ai #mlops #aiinfrastructure #ml #llms

Interesting finding on application dependencies from Sonatype State of Open Source Report.

Most Pulled Ecosystems
- JS: 4.5 Trillion
- Python: 530 Billion
Outdated software continued to be pulled
- 13% of Log4J downloads are vulnerable
- 80% of Application dependencies are never upgraded

Language Ecosystems Coverage
- Only 10.5% of open-source components are actively used out of the 7 million available
- 180 is the average number of Open Source components per application

Vulnerability Remediation
- The report states the average fix times for even critical vulnerabilities is 200-250 days, with some taking over 500 days to fix.

What application frameworks are you mostly using in your environment? What makes it challenging for application development teams to keep all dependencies up to date during initial development and once in production?

#applicationsecurity #supplychainsecurity #applicationdevelopers

Thanks to AI, there is a fascinating shift in how information retrieval business models are about to change. Many SaaS applications focus on indexing/processing and making data available to users for many use cases.

As AI components are introduced into enterprise products, the COGS of information retrieval fundamentally changes. Before you could index/store and make it available with predictable processing, GenAI introduced even more variable compute costs that scale with usage more than other use cases. Just like offloading all your workflows to “the cloud” can add up quickly, seeing how inference costs behave overtly will be interesting. This ignores the training costs of custom models that compound this problem.

#ProductStrategy #ai