Pete Skomoroch
@peteskomoroch
Followers
52K
Following
94K
Media
3K
Statuses
69K
Investor and AI startup founder. Focus: AI, LLMs, LifeOps, AI Product Management. Was founder @SkipFlag. EIR @Accel. Data Science & ML @LinkedIn, @AOL & @MIT
San Francisco
Joined April 2008
Many companies struggle to get impact from machine learning projects, so @mikeloukides and I wrote this guide: "What you need to know about product management for AI"
14
134
407
@altryne I dislike the entire framing of “helpful assistants” in the LLM world. If a person talked like this you as a kid you would think they were creepy. The kids are properly detecting that.
3
14
2K
Stop what you’re doing and read this entire post on the GPT-4 code interpreter plugin. This is completely bonkers and is going to change everything.
43
312
2K
ML people: Do you have a tool in your current model training workflow that is a secret weapon? Maybe something that labels data automatically, autotunes hyper-parameters, simplifies testing or does feature extraction? Tell us your secrets :).
114
267
2K
MIT reinstating standardized test requirement for applications. Not having SATs/ACT scores to consider tends to raise socioeconomic barriers and using tests like the SAT improves diversity in admissions:
71
374
2K
This open letter on math education in the US urges @CADeptEd to reject the latest draft California Math Framework (CMF) which eliminates 8th grade Algebra I. It has over 150 signatories, including Fields Medalists, Turing Award winners, & Nobel laureates:
63
356
2K
🤯New paper: “Self-Refine: Iterative Refinement with Self-Feedback” shows LLMs can improve themselves without humans. “SELF-REFINE is unique in that it operates within a single LLM, requiring neither additional training data nor reinforcement learning.”
35
236
1K
Video of SF School Board renaming of Sanchez Elementary: "Ok, so Sanchez is um, again here's another, this is another um colonizer who um basically". "Colonizer, California Missions, bla bla bla bla bla, it's in the notes" [marks Yes: rename]. They researched the wrong Sanchez.
83
238
1K
Using PyTorch + NumPy? You're making a mistake. A bug that plagues thousands of open-source ML projects:
14
206
1K
GPT-4 is one of those rare “drop everything” moments where science suddenly takes a giant leap forward. I don’t think most people outside of machine learning understand how significant this is.
Best thing I’ve read on GPT-4’s capabilities. You should read it. Impressive qualitative jump over ChatGPT. It’s definitely not just memorizing, it’s learning to think and reason. Probably the most important thing happening in the world right now. Thread with some highlights:.
31
106
1K
This Google paper basically uses ChatGPT to create a primitive 25 person v1.0 of “The Matrix”
Generative Agents: Interactive Simulacra of Human Behavior. abs: .project page:
21
128
899
GPT-4 appears to do a kind of “social mirroring”. If you give it an intelligent question about an unsolved problem a PhD in computational neuroscience might ask, you get a more intelligent answer. A similar question in the style of a 5th grader gets a 5th grade level answer.
68
69
896
Me: “I need to get a coffee”.My 7yo: “You don’t *need* a coffee, you *want* a coffee. Years ago people didn’t even have coffee. You’re addicted to it.”.
29
54
837
Databricks just released Dolly 2.0: the first open-source, instruction-following LLM licensed for commercial use. - 12B parameter LLM based on EleutherAI pythia model family.- training code, dataset, and model weights all licensed for commercial use.
20
165
817
Has anyone built a machine learning model to predict whether an avocado is ripe? 🥑.
70
98
669
@Reza_Zadeh It's a small step from dancing to martial arts. Once one robot knows kung fu, they can all know kung fu.
8
11
597
Too many years of machine learning made me misinterpret this tweet:.
15
94
656
PolyAI is releasing a collection of conversational datasets consisting of hundreds of millions of examples
0
202
616
Why is someone who doesn't have an undergraduate degree in mathematics the lead author of California's new K-12 mathematics framework?.
62
57
543
“I uploaded a SQLite database to Code Interpreter and it did everything on my roadmap for the next two years.”
@codethazine @natfriedman You definitely have to change your workflow to make the most out of Code Interpreter. I talk about how I use it here (click "41:14" for the audio) Here's how I use it to run micro-benchmarks against Python and SQLite:
9
65
557
Stumbled on this interesting @Wikipedia page today: "List of datasets for machine-learning research"
4
111
538
As a rule of thumb, you can expect the transition of your enterprise company to machine learning will be about 100x harder than your transition to mobile.
14
161
527
I’m consistently amazed at the lengths people go to avoid relational databases, good schema design and ETL, which would solve most of their needs
15
154
489
If so many applicants have near perfect SAT scores, we should make the test more difficult instead of not using it in admissions.
Fact check: In a recent admissions cycle there were:.2,000 available slots at Harvard College.8,000 U.S. applicants with perfect GPAs.4,000+ applicants ranked 1st in their high school classes.18,000+ scored 700+ on SAT reading/writing.20,000+ scored 700+ on SAT math.
40
28
494
Me just now: “He quit Google and moved to a farm in Australia where he works on robots and writes algorithms to count bees”.
16
17
470
What are the most popular machine learning podcasts right now?.
36
52
448
This paper uses a new prompting technique called “expert prompting” that generalizes the “imagine you are Jeff Dean” approach to writing prompts. 1) prompt the LLM to suggest named experts who could answer a given question. 2) have the LLM imagine the answers those named experts
14
64
439
This is quite a take from Judge Michael Klaeren on the 3 men who plotted to kidnap Gov. Gretchen Whitmer:. "Klaeren added that an encrypted communication network, not accessible to the general public, was 'in many respects no different than thinking the thought to yourself.'".
NEW: A judge has dismissed threat of terrorism charges against three men accused in a plot to kidnap Michigan Gov. Gretchen Whitmer . "There has to be some form of intent here to incite mayhem," the judge said.
48
150
366
Product Managers who can write decent specs and work with GPT-4 now have infinite engineering headcount.
have a graveyard of tightly scoped side projects that i’m just feeding to this thing and it’s just spitting out code…….that just executes lmaooo
3
45
412
If you live in San Francisco and stand in long lines at the pharmacy waiting for medication this Thanksgiving, you can thank @DeanPreston. His policies encouraged shoplifting, closing of @Walgreens locations, and limited access for seniors. Pharmacy line is out the door at 3pm:
57
55
383
Do you do machine learning consulting? Add yourself or your company to this list of ML consultants and I'll send relevant work your way:
12
104
391
The SF School Board will officially rescind their resolution renaming 44 schools, many via a blue ribbon panel researching the wrong historical figure or other factual mistakes. No members of the previous panel should be included in any future advisory panels on school renaming.
SF School Board has finally capitulated and will repeal its resolution renaming 44 SF public schools. Like a bully confronted in the school yard, they call our case frivolous while walking away. Huge thanks to Professor Tribe for standing beside us.
13
55
371
“Backspace is All You Need”. The LLM training technique in this Stanford paper helps avoid compounding errors in text generation by introducing a backspace token during training:
Introducing *SequenceMatch*, training LLMs with an imitation learning loss. Avoids compounding error in generation by:.1. Training against *different divergences* like χ^2 with more support OOD.2. Adding a *backspace* action: model can correct errors!.1/7
11
67
363
SFUSD reports the student body at Sanchez Elementary is 81% Hispanic with many students whose primary language is Spanish. Sanchez is one of the few schools in SF with a name tied to Mexican heritage.
9
18
342
SF politicians continue to find new ways to hurt kids and education. Refusing a $25M donation to bring summer school to 20,000 kids is shameful.
In December, San Francisco condemned the naming of a hospital after Mark Zuckerberg, who gave it $75 million. Now, one San Francisco lawmaker has temporarily derailed a proposal that includes a $25 million donation to bring summer school to 20,000 kids.
18
82
336
This is madness. Anti-asian racism and antisemitism is rampant. University administrators have lost their minds:.
Despite earning 3.97 unweighted and 4.42 weighted GPA, scoring 1590 out of 1600 on the SAT's and founding his own e-signing startup RabbitSign in sophomore year, Zhong was rejected by 16 colleges he applied to.
19
57
348
2020 was mostly awful, but one positive thing I focused at home was getting healthy. Finally hit my weight goal today:
31
0
349
A Tale of Two Cities: . San Francisco, 2016: SF School Board eliminates 8th grade algebra from the curriculum, calling it "racist" Pasadena, 2016: PUSD adds an accelerated program where 8th graders can take Calculus.
22
67
327
Commissioner Lopez just rolled her eyes at a public caller and was called out on it, then she got angry and appeared to turn off her camera.
33
33
307
What are the most mathematically rigorous deep learning courses with full video lectures online?.
14
77
336
Amazon should pre-print #BlackLivesMatter on all their cardboard boxes for use as protest signs.
6
53
307
When your cookies are the answers to machine learning interview questions
7
60
309
I didn't think it was possible, but I became a morning person in 2020 and I'm finally getting good sleep. If you know me, that's a huge shift. Instead of going to bed at 3-4am, I'm getting 8 hours & waking up at 5-6am. Followed @JamesClear's Atomic Habits:
13
17
305
One of the most important projects in machine learning is @paperswithcode. Launched in 2018 to track ML papers with source code & the state of the art results, now it hosts:. • 3,990 benchmarks.• 2,009 tasks.• 3,224 datasets.• 41,871 papers with code .
3
70
298
Bay Area NIMBYs have achieved a new level of super villainy today by forcing @UCBerkeley to revoke admissions offers for 3000 freshmen. At what point will NIMBYs look at themselves in the mirror and realize they are the bad guys?
Due to a NIMBY CEQA lawsuit to stop UC Berkeley from expanding to meet growing student need, the University is now forced to rescind 1000s of admission offers. Let’s be clear: This was never the point of CEQA. This broken status quo must change. Stay tuned
19
54
284
What if ChatGPT making small mistakes is a feature and not a bug? . Writing code with GPT-4 is addictive. Code it writes in response to prompts often “almost works”, so you iterate until it finally works and you “win”. This is how slot machines work: intermittent variable reward.
This is the thing I find most interesting about ChatGPT as a learning assistant: it turns out having a teacher that's mostly right but occasionally very wrong makes me think much more critically about what I'm learning, which I think is helping me build more robust mental models.
28
44
294
The unfortunate truth of the SF School Board recall is that the left should have handled it internally before it got to this point. SF Mayor, State Senators, Board of Supervisors, etc all asked them to resign. Imagine the ego to not quit, but instead sue your coworkers for $87M.
3
18
289
Dropped my daughter off for her first day of non-Zoom 4th grade this morning. The last time she was in a classroom for in-person instruction was 413 days ago:
10
8
277
Why Data Is Never Raw: “The way data is collected places limits on the inferences we can obtain.”
4
109
271
California is proposing not allowing Calculus in high school (which is absolutely needed for STEM majors), while inserting some nonsense “data science” course curriculum in the name of “equity”?.
This attack on opportunities for public school students to take advanced math is the result of a once every 7 years review of math curriculum. I encourage you to read the 2021 proposal for yourself. Then submit comments to .mathframework@cde.ca.gov
39
70
264
The real billionaire flex is to pay for a fleet of dozens of private jets, and keep them in the air continuously, carrying actors that look like you to random cities 24 hours a day. This confuses enemies trying to track your whereabouts.
32
11
265
Does this exist? . A chrome plugin powered by GPT-4 that reads the 30 open browser tabs you opened while researching a topic and synthesizes all the information into a single readable (and cited) report.
34
14
269
Why do people in in SF government lie about things that are easily checked? So bizzare.
46
9
251
For people who don’t understand why this is really happening in SF: there is an entire industry of nonprofit grifters who import drug users in order to collect government payments for housing and services. Enabled by City Supervisors like @DeanPreston and @ACLU
Stay focus pay very close attention to this video and you will see why and how it’s easy for people to stay addicted on San Francisco Streets This was filmed at 9:20 this morning on McAllister behind the Civic Center
9
36
255
"Big companies should avoid building their own machine learning infrastructure. Almost every tech company I talk to is building their own custom machine learning stack and has a team that’s way too excited about doing this." - @l2k dropping ML knowledge
3
75
242
Excellent first-hand account of one person's productivity gains from LLMs:
9
43
258
No-code taken to the extreme is a time bomb of complexity that results in an army of people maintaining a point and click legacy mess while gaining non-transferable skills.
justin's rule: all no-code eventually converges to yes-code over time. basically, at some point the UI becomes so unwieldy and hard to use that SQL is a superior UX.
7
38
242
This is ridiculous and offensive. Math education in San Francisco public schools is self-destructing and @SFUSDMath is proposing 3 years of meetings instead of fixing it. Voters demand a normal K-8 math curriculum update immediately that includes 8th grade algebra for Fall 2023:.
.@SFUnified, San Francisco's kids cannot wait. We don't need *THREE YEARS* to audit, evaluate, and pilot before implementing changes to the K-8 math program. Cut this timeline in half.
26
47
219
I've been trying to understand how @Scott_Wiener could screw up SB 1047 so badly, then I saw this Twitter thread. It looks like some of his top AI advisors and co-sponsors for this bill are two kids, and one of them just graduated high school? . It looks like they run an org
.@Scott_Wiener is presenting his AI framework #SB1047 in the Asm privacy committee. Says the bill will NOT hurt start ups. “There's been some inaccurate information and scare tactics online.”
26
29
230
My 8 year old gives me painfully accurate performance reviews now: “Daddy, sometimes when you explain something you talk too long and it gets boring. Can you write it down instead and I’ll read it later?”.
14
15
227
"We've developed a new revolutionary deep-learning program aimed at your kindergartner called Keras4Kindergartners"
10
68
222
The Current State of Machine Intelligence: great synthesis of the ML landscape by @shivon
http://t.co/UBtlR2PKvb.
3
195
228
I’ve talked to otherwise smart people who dismiss GPT-4 because they tried it briefly with vague prompts and it gave them uninteresting responses. They appear to be asking the wrong questions.
9
3
225
Apologies to followers outside of SF for the number of SF city government related tweets lately. It’s been difficult over the last year to see so many mistakes and denials of mistakes by the SF Board of Education, Board of Supervisors, and now the DA.
13
5
213
Idea: @Google should repurpose the Google Business system that already uses mobile phone data to measure how busy businesses are for real time pandemic social distance measurement
11
37
208
"I don't care about a plan." - Alison Collins last June right before voting against SFUSD Superintendent Matthews proposal to hire an external leader to drive school reopening.
Board member Alison Collins also worried that the consultants would end up relying on unpaid Black and brown community members to do the brunt of the outreach work. Hiring the consultants is “basically recreating white supremacy,” she said. (3/6)
27
34
202
Why many machine learning projects fail to deliver results and how product management is adapting to AI: - @annkspencer of @DominoDataLab wrote up an in depth summary of my recent London @strataconf talk.
4
79
215
It’s a strange and unsettling experience to see so many people going about their normal routines as if nothing has changed. Meanwhile others like Andrew see the shockwave starting to hit the world and are adapting as fast as they can:.
I made tools for GPT-4 to use - it will hallucinate when working with molecules directly. I instruct it to rely on these tools. First it does literature searches using one of these tools on the target. 3/n
7
35
216
Imagine the benefits to humanity of making digital scans of all books and academic papers ever published available as training data to GPT-4 and other LLMs. This should be a national priority:.
@reconfigurthing @QuintinPope5 @norabelrose So books + academic papers + US court cases from the past 10 years is approximately 150x the size of the Pile, or enough to train a chinchilla optimal 22.5T parameter model.
27
25
208
That derivative AI painting that sold at auction for $430k is a good lesson for scientists that marketing, connections, & PR strategy matters more than you think. This isn’t a meritocracy.
7
53
202
A woman did her aerobics class without realizing that the coup was taking place in Myanmar:.
Una mujer hizo su clase de aerobic sin darse cuenta de que estaban dando el golpe de Estado en Myanmar. Y pues puede verse como el convoy de militares llega al parlamento.
8
42
198
Jo Boaler claims she is devastated and only reached out to invite @minilek to a “chat”. Does this look like an invitation to a friendly chat?
Two professors at rival Bay Area colleges — UC Berkeley and Stanford— landed in a social media feud this week, trading accusations of racism and harassment. The backdrop for the tussle: California’s math wars.
16
20
190
There is a live stream of this SF government press conference on Twitter. They are walking in and out of the new bathroom and getting toilet paper to hold up at the podium. Meanwhile, China is building massive bridges in less time than it takes for us to approve new toilets.
A functioning public toilet in SF is a historical feat on the order of the Apollo moon landing, so of course they're holding a press conference about it.
14
21
195
Video of @SFUnified Clarendon Elementary renaming. "Clarendon was impeached by the House of Commons for blatant violations of the "heebeus corpus" [SIC] or whatever, for having sent prisoners out of England to places like New Jersey & holding them there without benefit of trial."
39
36
198
This week San Francisco continued dismantling advanced math & science education in the name of “equity” by cutting AP programs, yet a typical STEM worker earns two-thirds more than those employed in other fields. Public school students here will be less likely to get these jobs:.
Girls and women are systematically tracked away from science and math throughout their education, limiting their access, preparation and opportunities to go into these fields as adults. #WomensHistoryMonth2022 Learn more:
9
37
196
Hire the intersection of AI/stats people who understand this joke:.
It depends on the domain. I asked DALL-E to draw a diagram of WWII airplanes highlighting where they need extra armor, and it did a great job
5
11
188
Math is important, and that’s why we’re all looking forward to the appointment of a new School Board Commissioner who understands how percentages work.
10
6
184
Sanchez Elementary School was named for Sanchez Street in San Francisco, which in turn was named after Francisco Sanchez
5
11
187
Startups should stop doing whiteboard coding interviews and instead ask candidates to try to reset a kid's Minecraft password on a Mac using a Microsoft login backed by a Github account with Google Authenticator. etc.
14
22
189
This explains why @stitchfix keeps sending me boxes with dead cats in them!.
Stitch Fix is using something called eigenvector decomposition, a concept from quantum mechanics, to tease apart the overlapping “notes” in an individual’s style. Using physics, the team can better understand the complexities of the clients’ style minds.
13
23
190
Today in 2020: Reverse waterfalls.
A severe weather warning for damaging winds and damaging surf is current for Sydney, Central Coast, Mid North Coast, Hunter and Illawarra areas. At the moment, wind gusts of 70km/h are producing several reverse waterfalls in the Royal National Park.
7
78
189
cc @TonyThurmond @GavinNewsom @tedlieu @Scott_Wiener please read this letter and urge @CADeptEd to replace the Math Curriculum Framework and Evaluation Criteria Committee (CFCC) and start the Framework draft process over with new authors.
2
17
186
This. Stemming, punctuation and stop word removal, lowercasing. all these things will hurt you in real world applications.
NLPers, stop removing stop words "just cuz". I repeated a text classification tutorial ( but skipped the 'remove stop words' section and got a 2.8% INCREASE in accuracy. Stop words can improve your outcomes in many cases.
8
38
194
1000 people on the SFUSD School Board Zoom call and they are dividing public comment into 20 minutes against Collins and 20 minutes in support of Collins. This inflates a sense of public support where there is none, and is an approach they have followed before.
10
22
170
If the California Mathematics Framework is not rejected by @SBECalifornia, K-12 education in CA will spiral into disaster. It’s shocking: there is no evidence to support this curriculum, and self-citing CMF author @joboaler is profiting from her gov role.
12
32
186
In addition to Algebra, we should probably be teaching kids SQL in middle school. It's a technical skill that is relatively easy to pick up, and knowing it is like a superpower that allows you to answer challenging questions with data:
15
11
187
Github list of #NIPS2017 resources, video and slides of all invited talks, tutorials and workshops
0
69
185
Learning loss & end of merit @ Lowell HS: “I have three times as many students as usual failing—instead of one or two, I have three to six,” Wenning, the biology teacher, told me. “I have some students who have done no work the whole first grading period.”
21
37
182
