Make big graphs less of a hairball in the latest Graphistry, Inc. pygraphistry with one line: `g.modularity_weighted_layout().plot()` !
Same-community nodes cluster together; reuse w/ existing layouts; opt-in GPU mode 🤠
Before/after on cleaning up a crypto money laundering investigation 👇
This was a fun one -- a surprisingly simple concept goes super far in practice, slotted nicely into the system, and a nice trick in practice before going into heavier methods like graph neural network embeddings
Founding vision of Graphistry. GPUs + program synthesis + data science, end-to-end, everywhere. We were early (=> Nvidia RAPIDS, Apache Arrow, GPU SQL, etc), and the rest of the world took a hard turn onto the new pipeline last year. Busy times!
What if the actual outcome of LLM, AI, chat... is that computing w/ GPUs is so thoroughly commoditized that it becomes mind bogglingly obvious to integrate deep learning with everything.
So long as the appetite for GPUs is insane, the pressure to drive down their costs will grow, more competitors will arise, and their ubiquity in everything will massively increase
It's not about changing where / how users interact with computing... its not about LLMs, AI, ?? per-se... its changing the entire artifice of computing to make GPUs as commoditized as CPUs.
Big things happening at Graphistry, Inc. / #LOUIE_AI as we hit yet another record year. So, exciting to say... three new positions opening 😎 Federal data science + data eng for crime-fighting, and cyber area lead.
Check our careers page. Each is high impact in different ways - we're still early enough that each person has a lot to own & lead in!
Powerful:
- Reasoning-grade LLMs now generate good answers faster than we read: 50-100 words/sec
- That's 5-10X over last year's GPT-4, and we can assume another 5-10X next year
- Agentic tools already run fast
Many implications fall out of this. For example, genAI-native is moving from meaning "streaming text UI" to supporting "instant agentic iterations".
Repeating "Why Now" for graph RAG / autoKG in my conversations with infra/AI/ops leaders the last few months, whether they're in enterprise/gov/tech, similar shifts here --
BEFORE => AFTER
primitive KG triples => linked rich text summaries, embeddings, & citations
naively chunked vector index => hierarchical summary vector indexing
manual rule-based indexing => autoextraction & autosummarization w/ customization hooks
vector vs KV / text search => hybrid search, parallel indexing + ranking
myopic one-shot k-nearest-neighbor RAG Q&A => summary indexing, agentic reasoning, etc
manual ontologies => LLM autoinference + optimal learning loops
batch RAG & summary indexing => real-time ingest, indexing, & eventing
---
Curious if other shifts folks are seeing here?
Likewise, we're gathering design partners as we work towards more comprehensive, reliable, & scalable forms of RAG. We're very much up for chatting on challenge problems here, lmk! (And for cyber folks facing real-time intel challenges, we'll be at blackhat/defcon in a couple weeks.)
The question of "do we hold LLMs to diff standards to people", especially for misinformation analysis, is useful when thinking of where LLMs fit in our fight here. We & our partners have been going in a few directions around genAI for misinfo because of that:
* LLMs can be thought of as an army of low-level anthropologists. We and our partners are working to map out - in real-time - the different strains of controversy flying around. Note that they are NOT classifying truths, but instead tracking strains of controversial narratives.
* LLMs can be thought of as low-level wiki editors. They are not arbitrating truth, but summarizing what is notable and controversial, and to whom.
* LLMs can be through of chief-of-staffs. When you have a question, they are not doing the base research, but given your preferences, gathering the relevant info and present it how you like. For example... is a headline based on non-consensus thinking that it traces to a controversial narrative.
Associate Partner at McKinsey & Company | AI Researcher | Keynote Speaker
LLMs have demonstrated incredible potential across various domains, offering substantial benefits in areas like natural language processing, translation, and even creative writing. Despite their impressive capabilities, one significant concern remains: hallucinations. Hallucinations occur when LLMs generate content that appears plausible but is factually incorrect, often due to their inherent design and the vast, sometimes inconsistent data they are trained on. This issue has raised questions about the reliability and safety of using LLMs for critical applications.
But who is more rational? Humans or LLMs?
Research from Stanford University and UC Berkeley highlights that humans often cling to pseudoscientific beliefs and conspiracy theories despite evidence to the contrary. This study compared the susceptibility of humans and LLMs to unwarranted beliefs using the Popular Epistemically Unwarranted Beliefs Inventory (PEUBI). This psychometric tool includes 36 questions covering areas such as pseudoscience and conspiracy theories. Researchers prompted LLMs like GPT-3.5, GPT-4, and Gemini with these questions and compared their responses to the average human responses from previous studies.
The study revealed that humans scored an average of 3.05 on believing in pseudoscience, indicating a moderate level of belief in unwarranted notions. In contrast, LLMs generally showed lower susceptibility to these beliefs. For instance, GPT-4 scored as low as 1 on certain conspiracy theories, such as “The United States government knew beforehand about the terrorist attacks on 11 September.” These results suggest that while LLMs can generate hallucinated information, they may still be more rational compared to the average human in certain contexts(e.g., "The United States government knew beforehand about the terrorist attacks on 11 September") (see here: https://lnkd.in/ecAcPqi7).
The comparison between LLMs and human belief formation becomes even more intriguing when considering misinformation detection. A study from the Illinois Institute of Technology explores how difficult it is to detect LLM-generated misinformation compared to human-written misinformation. The findings indicate that LLM-generated misinformation is often harder to spot due to its sophisticated and deceptive nature. For example, LLM-generated misinformation has a higher attacking success rate in bypassing detection systems, with rates as high as 100% for paraphrase and rewriting generation, compared to 5% for human-generated misinformation. This suggests that LLM-generated misinformation is more deceptive and challenging for both humans and automated detectors to identify (https://lnkd.in/eBbEgRNn)
The results from these studies raises an important question: Are we holding LLMs to a different standard than humans when it comes to the reliability of information?
Something I suspect may be harming folks getting advice from otherwise lauded & successful engineers is their current career is rooted in the last ~2 decades of FAANG/VC/etc growth experiences that are from a universe that no longer exists. They were "children of summer", and it is now the fall:
* The current wave of senior devs largely came to power off of a decade of Zero-Interest Rate Phenomena projects where anything with a semblance of a pulse got funded and grew. Blockchain, apps, data science, cloud, social, unnecessary NIH in infra projects, etc. Even when projects were boondoggles, bad engineers could job hop within 1-2 years for a significant promotion in title and salary. I saw folks with 'senior' in their title after 1-2 years!
* Now, management is more careful to check the profit and losses of a project and hiring is capped overall. That has created brutal competition in the few remaining growth areas like genAI, so no free lunch there either. Past growth job fields like data science, data engineering, and cloud engineering are much better understood, so poor practices in those previous growth areas no longer pass muster in most rich organizations - it used to be easy to pitch crappy projects that are basically random forest or lift-and-shift for many people+$$$$, but no longer.
Some people have had to 'earn it' -- truly non-traditional background outsiders, startup founders who bootstrapped and survived multiple near-death experiences, etc. But most haven't, and even those who did, should be honest that they were lucky to ride the ZIRP era's halo effects. Not all the advice is bad... but I recommend a giant dose of extra skepticism.
Notably, if someone is pushing the same book, blogposts, and concepts, it is fair to ask, "what parts are now baseless and must be fundamentally reapproached? What is still true, and more importantly, still the most critical?"... and the answer better be well-thought & compelling.
Good luck out there!
Engineering @ Meta | Writing About Software Engineering & Career Growth
Zach Wilson went from Junior (IC3) to Staff (IC6) in 5 years through several successful job hops (Meta → Netflix → Airbnb).
I interviewed him recently because I was curious how companies were willing to give him next-level Senior/Staff interviews. Here's a brief summary of his time in big tech:
Meta: Junior (IC3) → Mid-level (IC4)
• Promoted in one half due to excellent execution; speed of promotion likely because they determined he was hired at the wrong level.
• Got two “Greatly Exceeds Expectations” ratings yet wasn’t promoted to Senior (IC5). Switched managers multiple times during this process due to reorgs.
Netflix: Senior (IC5)
• His previous manager at Meta recruited him to join Netflix which only hired Senior engineers at the time. Zach interviewed well and was able to secure the job-hop promotion.
Airbnb: Staff (IC6)
• He negotiated well with Airbnb and received a Staff interview loop and offer. This was only possible because he had interest from Google, Meta and Airbnb at the same time.
It was an interesting conversation with several other takeaways, for more details you can see my full newsletter post here: https://lnkd.in/gWZxg9tn
Big week for the Louie.AI team and everyone in genAI: Fast-following OpenAI's gpt4o, Meta has open sourced the first reasoning-grade AI model. Beyond many of our expectations, the 400b model already matches OpenAI's latest release on core reasoning benchmarks like humaneval!
We're moving fast. I'm in DC this week working with many gov & partner teams, and it's part of every conversation. On one end, a lot on running the newer models FedRAMP'd in the cloud (Groq, Amazon Web Services (AWS)) and on-prem (ThunderCat Technology) for key missions. More important is the WHY. The 400b version is exciting for security-sensitive analysts using Louie.AI to interactively talk to their data, and for intelligence teams automating smarter reports/alerts/etc. The smaller edge versions haven't been discussed as much, but IMO, also exciting for better preprocessing of intensive data such as in our real-time genAI tech for continuous monitoring (ex: graph RAG for disaster intelligence & force protection with Disaster Tech). Counting the days till we deploy 400b w/ our first on-prem users!
Shout out to our friends at Meta doing the good work, including the trustworthy AI folks like Joshua Saxe: amazing work!
* Core evals: https://lnkd.in/gQHG8qu7
* CYBERSECEVAL 3: https://lnkd.in/gxJcUV7QSean G.Mary Grace "MG" KarchKurt SteegeJon StresingRobert SchumannBrad BebeeLipyeow Lim
Founder: Graphistry & Louie.AI. 100X investigations w/ genAI-native design and GPU graph AI. Hiring in senior cyber, let's talk!
3dLinks: * Pygraphistry (pip install graphistry): https://github.com/graphistry/pygraphistry * Notebook: https://github.com/graphistry/pygraphistry/blob/master/demos/more_examples/graphistry_features/layout_modularity_weighted.ipynb