Saturday, June 4, 2022

Facebook: AI and metaverse | 10 minutes to read

 小扎突然拆散AI部门,LeCun被迫投身元宇宙,Meta动荡高管持续走人 | Redian新闻

Here is the link. 

Meta高层动荡,还没有停…

整个AI部门曝出重组,就在二号人物桑德伯格(Sheryl Sandberg)宣布辞职后的第二天。

现在AI副总裁Jerome Pesenti,也宣布将在6月晚些时候离开。

同时组织架构层面的一系列变动也浮出水面。

其中最重磅的,原Facebook AI研究院FAIR将并入元宇宙核心部门Reality Labs。

同时,支持旗下各大APP的AI算法团队转移进产品工程团队,AI4AR团队加入XR硬件团队,“负责任人工智能组织”并入社会影响团队。

Image 1 



有意思的是,整个计划正是由即将离去的Pesenti亲自操刀。

他认为,过去集中的AI组织给Meta带来了大量外部影响力,但在技术与应用的深度整合上遇到阻碍。

新架构会把人工智能系统的所有权分配给各产品组,加速新技术在整个公司范围内的落地应用。

而AI副总裁这个位置也被他自己给改没了,过渡期完成后他就会离开Meta,下一步去向还未透露。

image 2



而一系列震荡里,大家最关心的还是FAIR…

image 3



首先,他确认了FAIR将成为Reality Labs研究部门(RLR)下属组织,受Michael Abrash领导,而AI只是RLR研究范围中的一部分。

其次,FAIR内部组织架构不变,仍由LeCun把握战略方向,与两位管理者Joelle Pineau和Antoine Bordes共同领导。

最后,FAIR这个名字现在有了新的含义,F不再代表Facebook,而是Fundamental。

合起来就是“基础AI研究院”。

image 4



那么,如今并入主攻AR/VR的实验室,研究方向会有所改变么?

LeCun给出的回应是,并不尽然。

Reality Labs现在已不仅限于VR,用一套公关辞藻来说:

image 5



实际上,别看Reality Labs名字听起来不太厉害只是个“实验室”,但其实已经成为Meta的元宇宙核心。

Reality Labs最早由Oculus VR设备业务发展而来,后来又增加了智能眼镜、混合现实头盔等硬件产品及企业解决方案。

现在更是负责实现扎克伯格元宇宙愿景的重要部门。

据The Verge日前一篇报道,Reality Labs在短短一年时间里增长了约7000名员工,总员工已超过17000人。

人数上,已占整个公司的五分之一。

铁了心走元宇宙路线的Meta并非一帆风顺。自2月份年度财报发布,市值一夜蒸发2513亿美元以来,到现在股价也没缓过劲。

Reality Labs年度亏损高达100亿美元,最新的Q2财报显示,其成本同比又增长了55%,高于收入的增长速度,意味着亏损今年还将扩大。

最近,Meta现任首席技术官Andrew Bosworth也透露,一些Reality Labs的项目正在缩减或推迟,并停止招聘某些职位。

进入2022年以来,Meta的AI人才也在流失。

包括带领过强化学习研究的Edward Grefenstette,AI研究工程经理Heinrich Kuttler等至少4位知名AI研究者离职。甚至位于伦敦的AI实验室失去了大部分顶级研究者。

当时外界对这波离职潮的原因并不太了解,如今看来,可能正是这次组织架构调整的影响。

今后FAIR成为Reality Labs旗下一个部门,实际会受到怎样的待遇,还是个未知数。

All in元宇宙以来,至少20+高管离职

事实上,FAIR这一系列变动,还只是整个Meta震荡的“冰山一角”。

自扎克伯格宣称要All in元宇宙以来,Meta已经有20+核心高管和顶尖AI科学家离职。

在这些“出走”的高管中,从CTO到各部门负责人,有不少都是在Meta干了5年多、甚至10年往上的。

据新浪科技统计,仅在官宣改名Meta的2021年,离职的核心高管就有18名。

一方面,不少领域的技术高管都已经先后离职。

变动最大的要属已经Mike Schroepfer,离职前曾经担任了近9年的公司CTO。

他在任期间,曾经领导Facebook解决平台虚假和黄赌毒等内容的泛滥,期间Facebook也曾发展过一个叫做虚拟助手Facebook M的产品(类似Siri),但后来这一项目却因为各种原因被砍掉。

相对地,之前负责元宇宙相关项目的部门主管Andrew Bosworth则取代他成为新任CTO。

但其实,不少虚拟现实、AR/VR相关项目的负责人也并未久留。

包括Reality Labs商务合作副总裁Hugo Barra、AR/VR内容副总裁Mike Verdu、Facebook AI产品主管Ragavan Srinivasan等人,都已经于去年辞职。

去年是这样,今年也同样如此。

今年1月左右,不少Meta AI科学家纷纷离职,有些加入了谷歌,有些则加入了像Hugging Face这样的初创公司。

3月,Meta AR眼镜负责人Nikhil Chandhok宣布离职;4月,CNBC消息称Meta AI又有至少4名核心AI技术骨干离开,其中有不少都是在顶刊/顶会上发表过数十篇论文的大牛。

另一方面,则是广告等核心业务的高管变动。

例如主要负责公司广告业务、监督销售和营销部门的首席营收官(CRO)David Fischer,以及全球广告销售副总裁Carolyn Everson,都已经于去年离职。

此外,包括Facebook App负责人Fidji Simo和商务副总裁Deborah Liu、Facebook数字钱包负责人David Marcus和副总裁Kevin Weil、产品副总裁Asha Sharma等人,也都已经离开。

就在昨天,Meta二号人物Sheryl Sandberg官宣辞去COO职位,无疑将这一波“离职浪潮”推向了顶峰。

image 6



此前,Sandberg在公司的地位举重若轻,甚至有人形容她是“影子CEO”。

Sandberg在任期间,一举带领Meta旗下的广告业务从0增长至如今的1150亿美元年收入,而这也是Meta的主要收入来源之一。

如今广告收入仍然占Meta总收入的97%,但这一业务增长情况正逐年下滑。

值得一提的是,在Sandberg这篇“千字告别信”中,也只字未提对元宇宙的展望或看法,只是回顾了自己过去十几年来在Meta的工作感受。

至于这一波变化浪潮过去后,Meta究竟能否在元宇宙的概念中生存下来,还得交由时间判断。

One More Thing

事实上,FAIR也曾经是个产出了大量前沿研究和一批大牛的“明星AI实验室”。

自2013年LeCun组建以来,FAIR不仅在Meta内部有大量影响力,更对外开发了大量基础研究成果、代码和数据集,成为业界一个传奇组织。

研究人才上, 除了领头人三巨头之一LeCun,还有田渊栋、何恺明等业界大牛都曾在这里做出不少顶会级研究成果。

深度学习框架上,从最早的Torch到PyTorch,已成为整个生态内、特别是学术界的主导。

工业算法上,Mask R-CNN及其后续改进变体一度成为业界实例分割的主流方案。

开源工具上,目标检测平台Detectron,序列建模工具集Fairseq都是GitHub上最受欢的项目之一。

前沿研究上,自监督学习一直是FAIR最重视的研究方向,从MaskFeat到最近的MAE,为CV领域的自监督学习铺开了新的道路。

Sandberg could be running for office, says top Meta watcher

June 4, 2022

Here is the link. 

Gene Munster, Loup Ventures, reacts to news that Sheryl Sandberg is stepping down as Meta's COO. With CNBC's Melissa Lee and the Fast Money traders, Tim Seymour, Bonawyn Eison, Karen Finerman and Steve Grasso. For access to live and exclusive video from CNBC subscribe to CNBC PRO: https://cnb.cx/2NGeIvi


Friday, June 3, 2022

Leetcode discuss: 2263. Make Array Non-decreasing or Non-increasing

Here is the link. 

C# | Quick learner | Maximum heap | Sorted<Tuple<int, int>> | Tip: Add twice

June 3, 2022
I like to be a quick learner. I could not figure out the trick by reading C++, Java code shared by votrubac. So I debugged and then figured out how to make the test case work.

Test case | {3, 2, 4, 5, 0} | Trick: Add twice into the max heap
I like to go over the test case {3, 2, 4, 5, 0} and explain why it takes 4 operations to make a non-decreasing order in reverse order.

I chose to use C# SortedSet<Tuple<int, int>> to simulate a maximum heap.
I tried to work on the above case step by step by myself. I am working on

cost(nums, -1)

First, last element in the array, index = 4, add it to Max heap, heap - new Tuple<int, int>, (0, 0)
Next, index = 3, value 5, add it to heap, (5, 1)
Next, index = 2, value 4, heap.Max.Item1 = 5 > 4, so result variable increments (5 - 4) = 1, remove (5, 1) from heap,
add (4, 2) to heap <- inside if statement

if (heap.Count > 0 && heap.Max.Item1 > current)
{
	result += heap.Max.Item1 - current;
    heap.Remove(heap.Max);
    heap.Add(new Tuple<int, int>(current, count++)); 
}

Next, it is to add (4, 3) to the heap by executing the statement outside if statement

// explain: why {3, 2, 4, 5, 0} reverse order, 4 will be added twice? 
heap.Add(new Tuple<int, int>(current, count++));

Mext. it is to add (2, 4) to the heap, and result += (4 - 2), result = 3, remove (4, 3) from heap
Last, it is to add (3, 5) to the heap, and result += (4 - 3), result = 4, remove (4, 2) from heap.

votrubac | Analysis
I just quickly copied the idea and analysis from votrubac in the following:

The greedy logic is quite tricky. We process numbers left-to-right, and put them into a max heap.

When the current number n is smaller than the largest so far m, we know that we need to do m - n adjustments.

Now, the tricky part. We can increase n and/or decrease m so that they become the same. But for the purpose of the greedy algorithm, we assume we decrease m all the way to be the same as n, and put the new value to the heap.

The following C# code passes online judge.

using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
using System.Text;
using System.Threading.Tasks;

namespace _2263_make_array_non_decreasing
{
    class Program
    {
        static void Main(string[] args)
        {
            var result = ConvertArray(new int[]{3, 2, 4, 5, 0});
            Debug.Assert(result == 4); 
        }

        /// <summary>
        /// study code
        /// https://leetcode.com/problems/make-array-non-decreasing-or-non-increasing/discuss/2011905/Greedy
        /// </summary>
        /// <param name="nums"></param>
        /// <returns></returns>
        public static int ConvertArray(int[] nums)
        {
            return Math.Min(cost(nums), cost(nums, -1)); 
        }

        /// <summary>
        /// checklist:
        /// 1. Maximum heap
        /// 2. Tricky part - greedy - decrease m all the way to be the same as n, and put the new value to the heap
        /// </summary>
        /// <param name="nums"></param>
        /// <param name="d"></param>
        /// <returns></returns>
        private static int cost(int[] nums, int d = 1)
        {
            int result = 0;
            var heap = new SortedSet<Tuple<int, int>>();

            var length = nums.Length;
            var count = 0; 
            for (int i = (d == 1 ? 0 : length - 1); i < length && i >= 0; i += d)
            {
                var current = nums[i];

                if (heap.Count > 0 && heap.Max.Item1 > current)
                {
                    result += heap.Max.Item1 - current;
                    heap.Remove(heap.Max);
                    heap.Add(new Tuple<int, int>(current, count++)); 
                }

				// explain: why {3, 2, 4, 5, 0} reverse order, 4 will be added twice? 
                heap.Add(new Tuple<int, int>(current, count++));
            }

            return result; 
        }
    }
}

SABR | Yahoo -> Finance

 Over the past three months, shares of Sabre (NASDAQ: SABR) moved lower by 15.47%. Before we understand the importance of debt, let us look at how much debt Sabre has.

Sabre's Debt

According to the Sabre’s most recent balance sheet as reported on August 10, 2020, total debt is at $4.69 billion, with $4.61 billion in long-term debt and $77.88 million in current debt. Adjusting for $1.31 billion in cash-equivalents, the company has a net debt of $3.38 billion.

Let's define some of the terms we used in the paragraph above. Current debt is the portion of a company's debt which is due within 1 year, while long-term debt is the portion due in more than 1 year. Cash equivalents include cash and any liquid securities with maturity periods of 90 days or less. Total debt equals current debt plus long-term debt minus cash equivalents.

Investors look at the debt-ratio to understand how much financial leverage a company has. Sabre has $6.13 billion in total assets, therefore making the debt-ratio 0.76. Generally speaking, a debt-ratio more than one means that a large portion of debt is funded by assets. As the debt-ratio increases, so does the risk of defaulting on loans, if interest rates were to increase. Different industries have different thresholds of tolerance for debt-ratios. A debt ratio of 40% might be higher for one industry and normal for another.

Why Investors Look At Debt?

Debt is an important factor in the capital structure of a company, and can help it attain growth. Debt usually has a relatively lower financing cost than equity, which makes it an attractive option for executives.

However, interest-payment obligations can have an adverse impact on the cash-flow of the company. Having financial leverage also allows companies to use additional capital for business operations, allowing equity owners to retain excess profit, generated by the debt capital.


SABR | SeekingAlpha

 

Sabre: No Valuation Room Left For H1 2021

Summary

  • Sabre has raised an additional $1.5B in capital during the pandemic, causing a significantly high leverage ratio.
  • The company will experience revenue recovery benefit stemming from the vaccine distribution, but this will take 2-3 years.
  • Meanwhile, the high debt level and high interest payment will pressure the operating cash flow and its equity value.
  • At its current price range of $16-$17, I no longer see an upside value in Sabre stock for 2021.

Thesis

While Sabre (NASDAQ:SABR) will experience revenue recovery from the vaccine distribution, the massive amount of debt and higher interest payment will pressure down the financial value of the company. At the price range of $16-17, I believe that the recovery of 2021 to 2023 is fully priced. The stock no longer provides financial value until the revenue recovery is fully materialized and the debt is paid off.

Sabre and the legacy of the 2020 pandemic

Sabre is a global distribution systems (GDS) provider for airlines, hotel, car rental and other travel services. The GDS is an oligopoly industry, with the top three players(Amadeus (OTCPK:AMADY) and Travelport) dominating more than 90% of the market share. There is a high entry barrier to become a major player, as it is difficult for small players to obtain large scale ticket inventories from multiple airlines and to build a network with online travel agencies like Expedia (EXPE) and Booking Holdings (BKNG).

More than 90% of the company’s revenue (Travel Solution Revenue) is directly linked to the number of travel bookings. This makes the company’s business very cyclical to the travel volume and vulnerable to sudden shocks like the 2020 pandemic.

During the initial pandemic shock to the travel industry, the company’s revenue declined by -67% YoY during 2020. The company started to burn cash by recording a negative operating cash flow of -$773M during 2020.

To manage this unprecedented crisis, the company raised its capital by $1.5B. ($1.1B from loan, $375M from equity and $69M from the leaseback of their own building.) Also, the company has made its business model lean, via restructuring and reducing its global workforce by 20%. The company could increase the operating margin by migrating over from a legacy server to a cloud infrastructure. The company now expects to have $175M to $200M in annual savings from 2021 onwards, based on their last quarterly review.

Time to talk about the travel recovery speed and debt

There is no doubt that Sabre will survive the crisis as its business is essential for the online booking process, and it will have a strong business in the longer term due to the restructuring efforts. The recent stock price rally reflects this aspect.

However, we have to acknowledge that the travel industry recovery will take time. McKinsey report has forecasted that global air-travel demand will not recover until the end of 2024, which is corresponding to the IATA’s forecast. I will not go into the details of those forecasts, but it is reasonable to assume that the travel industry will recover to the 2019 level by the end of 2023 or 2024.

Over the next 2-3 years of recovery, the company will not lower its debt level significantly because it needs to keep its cash reserve until its operating cash flow recovery is strong enough. According to my forecast (see the last table at the end of the post), the company cannot generate meaningful cash flow that can pay off the debt of $500M - $600M until the year 2023. This is also evident in the company's own debt repayment schedule in the table below. According to this schedule, the major repayment will start from 2023 with $645M, followed by $1.9B in 2024.

This high debt level entails additional interest costs which ranged from $80M to $100M annually, compared to 2019. The company has paid $156M of interest payment in 2019, and this is increased to $235M during 2020.

Thus the recovery of operating cash flow will be offset by this interest payment from 2021 to 2024. Prior to the pandemic, the company generated an operating cash flow of $735M excluding the interest payments of $156M. I expect that the company will generate an operating cash flow before interest payment of $10M to $200M in 2021. This level of operating cash flow is not enough to cover annual interest costs of $230M. From 2022 to 2023, the company will begin to record the positive operating cash flow, but more than half of those values will be used for paying the interests. Of course, this means that equity investors cannot expect the dividends for at least 2-3 years.

Valuation and Conclusion

I used the DCF method to calculate the enterprise value and reflect the debt and cash balance to determine the equity value. The good news is that the company is not expected to have any further liquidity issue and additional capital raise because of its high cash balance of $1.5B.

The net equity book value may fall into negative territory for the first half of 2021. This does not mean that the stock price should be scrapped to $0, as the equity market value is the discount of future cash flow. However, the negative equity book value means that the company has a long way to fix its balance sheet and recover the financial value to equity holders.

My valuation for 2021 suggests that $16-$17 is in the expensive range with the assumption of fast travel industry recovery. So my conclusion is that the investors will have to be very patient on this stock if the expectation is for the price to recover to the pre-Covid level ($20-$22). I am no longer bullish on this stock for 2021 as the upside will meet the ceiling of the $17 range. I would come back to Sabre, once they can use strong operating cash flow to pay off the debt.

Daniel Cho profile picture
Daniel Cho
06 May 2021, 1:56 PM
@John Gilluly Normalized EBITDA $600M + cloud / restructuring effect $200M gives EBITDA 800M. If this is permanent effect, that gives 25% higher valuation. If I calculate low 2019 $20 * (1.25) = $25, or 2019 average value $22 *0.85 (Dilution effect) * 1.25 = $24...DCF results are similar that 2023 ~ 2024 the company value is $20 to $25.

But unfortunately, this is not going to happen until 2024...or maybe if market reflects this early in 2023, the price may go up by that time.

I think you had similar results with me before? just long time waiting...with travel recovery and debt problem.

I consider holding this stock in a long term even after that 3 years. If the air travel industry CAGR is more than 3% and SABR has more revenue growth than average with leveraing to its technology, I feel that you can stay with the company long time.

But I think how the new tech will impact the business dynamics is unsure...will keep monitoring it.
Oracle of Yomama
06 May 2021, 8:30 AM
I think the 2021 price target calculation understates the impact of the cost savings at a 2.5x multiple -- the capitalized value of $200M of annual cost savings ($144M net of tax at an assumed 28% rate) using your 9.5% DCF discount rate is $1.5B, so you would have to add that $1.5B to the 2019 enterprise value rather than $500M. Dividing $1.5B by the current 319M shares is an additional $4.75 of value per share versus the $1.57 of value that would be implied by a $500M value for the cost savings ( = ~$3 more upside).

2263. Make Array Non-decreasing or Non-increasing | Hard level

 Example 1:

Input: nums = [3,2,4,5,0]
Output: 4
Explanation:
One possible way to turn nums into non-increasing order is to:
- Add 1 to nums[1] once so that it becomes 3.
- Subtract 1 from nums[2] once so it becomes 3.
- Subtract 1 from nums[3] twice so it becomes 3.
After doing the 4 operations, nums becomes [3,3,3,3,0] which is in non-increasing order.
Note that it is also possible to turn nums into [4,4,4,4,0] in 4 operations.
It can be proven that 4 is the minimum number of operations needed.

15 minutes thinking

  1. Non-decreasing or non-increasing? 
  2. +1 or -1
  3. How to find minimum number? Brute force? 
  4. If the minimum number is 4, how to find the first one - increase one or decrease one? 
  5. Data structure? 
  6. Algorithm? 
  7. DFS? 
  8. Work on example: [3,2,4,5,0]

SABR

 A month has gone by since the last earnings report for Sabre (SABR). Shares have lost about 21.1% in that time frame, underperforming the S&P 500.

Will the recent negative trend continue leading up to its next earnings release, or is Sabre due for a breakout? Before we dive into how investors and analysts have reacted as of late, let's take a quick look at its most recent earnings report in order to get a better handle on the important catalysts.

Sabre Q1 Loss Narrower Than Expected, Sales Up Y/Y

Sabre reported a narrower-than-expected loss for first-quarter 2022. The company’s adjusted loss per share of 29 cents versus the Zacks Consensus Estimate of a loss of 38 cents. The figure was also narrower than the year-ago quarter’s 72 cents per share.

Sabre reported revenues of $584.9 million for the first quarter, which is significantly higher than the $327 million posted in the year-ago period and surpassed the Zacks Consensus Estimate of $564.2 million. This year-over-year surge in the top line reflects a gradual recovery in global air, hotel and other bookings.

Top-Line Details

The Travel Solutions segment’s revenues increased to $534 million from the year-ago quarter’s $288.9 million, primarily on the gradual recovery in global air and other bookings.

Distribution (sub-division of Travel Solutions) revenues improved to $342.9 million from $151.8 million in first-quarter 2021, chiefly driven by the gradual recovery in bookings and an increase in average booking fee due to a shift in bookings mix.

IT Solution (sub-division of Travel Solutions) revenues were $191.1 million, up from the year-ago quarter’s $137.1 million. This increase was primarily driven by the gradual recovery in the passengers boarded. The number of boarded airline passengers — a key revenue metric for the IT Solutions division — increased to 129.2 million from 75.2 million in the prior-year quarter.

The Hospitality Solutions segment’s revenues totaled $56 million compared with the year-ago quarter’s $42.2 million. This upside was mainly fueled by the gradual recovery in central reservation system transactions and higher Digital Experience revenues. Central reservation system transactions rose to 23 million from 17.6 million in the year-ago quarter.

Sabre reported an adjusted operating loss of $29.2 million, significantly narrower than the operating loss of $166.5 million posted in the year-earlier period.

Adjusted EBITDA improved from a negative $110 million reported a year ago to $5.2 million. This improvement was driven by an increase in revenues and a decline in the provision for anticipated credit loss.

Balance Sheet and Cash Flow

Sabre exited the March-end quarter with cash, cash equivalents and restricted cash of $1.21 billion compared with the previous quarter’s $999.4 million.

During the first quarter, the company utilized $139.1 million of cash for operational activities and generated a negative free cash flow of $156.5 million.

How Have Estimates Been Moving Since Then?

It turns out, estimates review flatlined during the past month.

The consensus estimate has shifted -7.14% due to these changes.

VGM Scores

Currently, Sabre has an average Growth Score of C, though it is lagging a lot on the Momentum Score front with an F. Charting a somewhat similar path, the stock was allocated a grade of D on the value side, putting it in the bottom 40% for this investment strategy.

Overall, the stock has an aggregate VGM Score of D. If you aren't focused on one strategy, this score is the one you should be interested in.

Outlook